[EMAIL PROTECTED] (akhan) wrote in message news:<[EMAIL PROTECTED]>... > Thanks for your response. > > Could you give me some suggestions on how to test whether a model > fitted the dataset significantly? > > Even if model A fits the data significantly better than model B, we > still don't know at what extent model A fits the data. Maybe both A > and B are far away from the truth.
Oh, how I wish I knew the truth! Some things to consider when searching for an expression that might represent reality: 1. If there is an appropriate scientific model use it as your base function. 2. Be sure that whatever model you use is asymptotically correct. That is, extrapolation to limits like +/- infinity should agree with logic. Similarly, one should avoid un-natural poles, jumps, etc. 3. The fitting function used should also be well behaved locally; a 12 term polynomial can fit 12 data points exactly, BUT it'll probably wiggle like crazy between the fitted points. Does this make physical sense? That is, you must think about how smooth the true functional form should be. 4. Measurements should be randomly scattered about theoretical predictions; if there are long runs where all the errors are the same sign the model probably isn't exactly true, even if the residual is small. 5. Apply Occam's razor: simple is probably best. Given two models that cannot be clearly distinguished based on statistical tests (like an F-test comparing the residuals of two models), I'd certainly go with the model consistent with the above criteria. . . ================================================================= Instructions for joining and leaving this list, remarks about the problem of INAPPROPRIATE MESSAGES, and archives are available at: . http://jse.stat.ncsu.edu/ . =================================================================
