Hello, U�ytkownik nothanks <[EMAIL PROTECTED]> w wiadomooci do grup dyskusyjnych napisa�:[EMAIL PROTECTED] > Hello sci.stat* readers, > > I have been working with SEMs (actually, the models I've been > working with are just Mixed models) and am curious why SEM > practitioners seem to assess model fit by only comparing the > observed and predicted covariance matrices. As opposed to (also) > using statistics based on observed vs. predicted outcomes. > > One can have an SEM with a covariance matrix that is fit very > well (or perfectly). Yet, if you compared the predicted and observed > outcomes, it's terrible! So, two things: > > 1. What is the rationale for ignoring this type of poorly fitting model > (seems, to me, to be a fundamental flaw - even if you are just using > if for "exploration")? > > 2. Is there software that gives you both kinds of fit statistics? > LISREL 8.3 and PROC CALIS seems to only give those based on > predicted and observed covariance matrix. And PROC MIXED only > allows certain paramaterizations. > > Thanks! > > (I am aware of SEMNET, but thought I would ask statisticians first.) > Ad. to campare pred. ver. outcom.: Don't want "to hurt" you, but nothing new to make "perfect" forecasting model. Could you tell: 1. In which period (ex ante or ex post veryfied in time) you model gives perfect forecasts? If it's ex ante (cut datas) be worry... 2. Could you tell something more about stability? En example: take regular harmonic regression (with n harmonic components) and check my sugestions - time to time better way is to take model with worse fit, but with ability to produce stability and probabilty of real forecasts. If you don't want to do this, take any simulation method (ex. neural network). In this case you will get PERFECT forecasts. Please, don't offend. I'm very interesting in you model...
With best regards, Oskar Czechowski . . ================================================================= Instructions for joining and leaving this list, remarks about the problem of INAPPROPRIATE MESSAGES, and archives are available at: . http://jse.stat.ncsu.edu/ . =================================================================
