Will Dwinnell wrote: > [EMAIL PROTECTED] (akhan) wrote > "Is there any statistical metod which can be applied to test whether a > non-linear model fit a dataset well or significantly?" > > One can assess how far off the model is expected to be from the > underlying process through any of a number of resampling procedures, > regardless of the modeling process (or even if you are reading Tarot > cards!). >
As well as the rest of Will's posting .... you need to clarify what you want to do ... If you want to "test whether a non-linear model fit a dataset well ", then there is the question of what information you have about measurement errors. A model which "fits a dataset well " should have residuals about the same size as the measurement errors and you can do a significance test for this if you have a useful source of information about the measurement errors. Your "experiment" might be of the repeated-measurement type, and this could provide the required info. If you want to "test whether a non-linear model fit a dataset significantly", presumably you mean "significantly better than model B", where model B is some simple model. The simplest case is where model B is a special case of your full model and you should usually be able to find a suitable simple model within the full model. A possible difficulty with non-linear models is that you may have for example single parameter for which a specific value will switch off the effects of a whole collection of others. Much of ordinary max-likelihood theory requires that a first-order expansion of the log-likelihood in terms of the parameters will be adequate in some sense, but this is not guaranteed. While some theory has been developed to cover more general cases, you probably need either to investigate your model for possible problems of this type before applying theoretical results, or else use a resampling or simulation approach. It is important to distinguish these two types of questions. David Jones . . ================================================================= Instructions for joining and leaving this list, remarks about the problem of INAPPROPRIATE MESSAGES, and archives are available at: . http://jse.stat.ncsu.edu/ . =================================================================
