(See David Jones' reply...) On Wed, 25 Jun 2003, praxis wrote:
> Assuming I do a multiple regression using ML estimation instead of > OLS, do I still need to meet all the assumptions like normal > distribution assumption, linearity assumption, and/or > homoscadesticity assumption? If yes, could anyone explain why? For the same reason(s) they were needed (or not, as the case may be) using OLS. WRT the three assumptions you have mentioned: Normal distribution. If you intend to test hypotheses about the values of parameters estimated in the regression analysis, SOME distributional assumption is necessary. (But make sure you know what this assumption means. A remarkably large proportion of people manage to make absolute hash out of what the assumption actually IS, let alone whether it is or is not violated in their data.) Linearity. If the relationship between the response variable and the predictor(s) is not linear, then you can't trust any of your output: essentially, you're modelling a curve of some kind as a straight line. (Or a more complicated surface by a plane or hyperplane.) This has to do with misspecifying the model. Homoscedasticity. This also has to do with possible misspecification of the model. If heteroscedastic, and the heteroscedasticity is systematic (e.g., small conditional variance for small values of X and large conditional variance for large values of X), this may be a signal that you've not correctly identified the response variable. (E.g., perhaps analyzing log(Y) instead of Y would be informative.) ----------------------------------------------------------------------- Donald F. Burrill [EMAIL PROTECTED] 56 Sebbins Pond Drive, Bedford, NH 03110 (603) 626-0816 . . ================================================================= Instructions for joining and leaving this list, remarks about the problem of INAPPROPRIATE MESSAGES, and archives are available at: . http://jse.stat.ncsu.edu/ . =================================================================
