[EMAIL PROTECTED] (akhan) wrote in message news:<[EMAIL PROTECTED]>... > Sorry for cross-posting. > > I'm trying to fit the sum of two or three normal distributions to the > frequency distributions of my datasets. And I used the Nonlinear Least > Square Fitting in OriginPro 7.0 to acquire the values of the > parameters of the two functions for each dataset. I want to know is > there any way to know whether one model fitted the data better than > the other one. Someone has used F-test to do this, but I failed to get > access to any literature associated with that. Can anyone tell me the > formulation of the F-test in this situdation? > > BTW, as the models (especially the three-peak one) didn't fit some > datas very well, and I got several sets of values of the parameters > for one dataset, can F-test be applied here to select the best set of > values of the parameters? > > Thanks a lot for your help!
A better way to fit a mixture of normals distribution is by maximum likelihood estimation via the EM (expectation maximization) or related algorithm. Two fairly recent books on the subject are Finite Mixture Models by Geoffrey McLachlan and David Peel The EM Algorithm and Extensions by Geoffrey J. McLachlan and Thriyambakam Krishnan The EMMIX program fits finite mixtures -- see http://www.maths.uq.edu.au/~gjm/emmix/emmix.html . When fitting a finite mixture model, a key question is the number of components. In the first book above, the authors say that information criteria such as AIC or BIC do a reasonable job. I suggest also computing the standard errors of your estimated mixture parameters and looking for degeneracy among the components. . . ================================================================= Instructions for joining and leaving this list, remarks about the problem of INAPPROPRIATE MESSAGES, and archives are available at: . http://jse.stat.ncsu.edu/ . =================================================================
