Dale, this discussion is getting off the subject raised by the original poster, but I will expand a little. Your suggestion is a good way to come up with variance or covariance estimates that are not too dependent on the model chosen. But these variance and covariance (depending on the model) are used directly in the formulas for the Wald (chi-square) or scaled Wald (F) statistics calculated in MIXED and GLIMMIX. The ANOVAF option in MIXED (not in GLIMMIX) does something different: it calculates a different test statistic (a different formula: see the MIXED User's Guide) based on whatever variances and covariances are estimated. That is, one gets the regular table of Wald test statistics, but also another table of the "Anova-Type-Statistics" (ATS) (with estimated df_N and df_D based on the variances and covariances). This test statistic also has an F distribution asymptotically (but it works great at small sample sizes). The ideas go back many years, and one can read some of this in the article by Brunner and Puri I mentioned (and the citations therein), as well as in the excellent book by Brunner et al. (Nonparametric Analysis of Longitudinal Data in Factorial Experiments). A more recent technical paper (on a more complex layout) by A. Bathke et al. (The Amer. Stat., vol 63, pages 239-246 [2009]) may be of interest. A less technical article for biologists is by D. Shah and L. Madden in the scientific journal Phytopathology (vol. 94, pages 33-43 [2004]).
For one-way layouts, there have been several tests for a while. Even in GLM procedure, the Welsh option on the MEANS statement gives one another test statistic that is more customized to unequal variances. The ANOVAF option in MIXED gives a different test, but is in the same spirit. The ANOVAF can deal with a wider range of models.
For the technical details, one does NOT need a subject option on the REPEATED statement for this situation (in MIXED), or for any situation where all the observations are independent. Your use this statement just to get a different variance (using group option) for each group. With correlated data (split plots and repeated measures), the statement is more complicated, and a group and a subject option are needed, with a type=un (usually).
The theory by Brunner et al., going back to the theory by Box (1954) is all moment based. Thus, one is really using the MIXED procedure simply as a crank to work through the calculations. (There are also macros available that do everything in IML,and also R scripts). Although I didn't list it, I should have put method=mivque0 on the MIXED statement to get moment-based estimates. It can blow up without this.
You might rightly point out that the Wald statistics perform pretty well with unequal variances. That is true, but with ordinal data, however, there can be many ties, and 0 variances for some groups. This really causes problems with the Wald statistics (things can go to infinity easily). But the ATS (ANOVAF option) is designed to work just fine under these circumstances (ill conditioned covariance matrix, even a singular covariance matrix).
Regarding your code with GLIMMIX, the empirical option won't give you separate variances for each group. Actually, your random _residual_ statement will result in two residual variances (the regular one and another one fixed at 1). You would need
random _residual_ / group=group subject=group;
to get what you want. And with the empirical option, you do need the subject= option .
... View more
