Not necessarily a contradiction. It is well known that individual confidence intervals can overlap (somewhat) when two means are significantly different, although many others are misled by the individual confidence intervals. This is true even with uncorrelated means (i.e., uncorrelated estimated expected values). Individual confidence intervals can have even more overlap when the means are correlated. By the nature of your model (with a random effect), there is the potential that the means are highly correlated. You didn't show all the output (such as the random effect variance), but I am guessing it is large (relatively speaking),which means there is high correlation in the means. For instance, if you square the SEs for each individual mean and for the difference, you have variances. V(1) is the square of the SE for mean 1, and so on. The standard formula for the variance of a difference (V(D)) is: V(D) = V(1) + V(2) - 2*COV(1,2) where COV(1,2) is the covariance of means 1 and 2. Using your results, you have: .01823^2 = .03846^2 + .03685^2 - 2*COV(1,2) .0003323 = .0014792 + .0013579 -2*COV(1,2) The covariance is not shown here. However, a little algebra shows that it would have to be 0.001252. You can use this to estimate the correlation of the two means. This would be: 0.001252/(0.03846*0.03685) = 0.88 . Thus, there is a high a correlation in your case. Bottom line, assuming you did other things correctly (can't tell from what you have given us), a consequence of your model fitted to the data is that the least squares means are highly correlated. Because of this, the standard error of a difference (SED) will be considerably smaller than the SED you would have with uncorrelated means. If the means were uncorrelated, COV(1,2)=0, and the SED would be: sqrt(.0014792 + .0013579) = 0.053. Your SED of 0.01823 is a lot smaller. I would definitely use the results of the differences of the LSMEANS. This is the test result that properly accounts for the correlation of the means (and the model fitted to the data). In fact, this is one example (assuming you did other things correctly) of the great advantage of using mixed models: you gain a great deal in precision in the comparison of LSMEANS. ... View more

It is not clear what you are asking. If you look at the MIXED output, you will see the pairwise comparison of all region*condition combinations (2*18=36 interaction means, which would give 36*35/2 = 630 pairwise comparisons). Your post-model fitting code is then eliminating all the pairs where the region is not the same. In other words, it looks like you will then print a table with 18 brain regions, and the estimate is the pairwise difference of the two treatment lsmeans. Is that what you want? Seems like a good plan to me. All the results in file DATAPet2 are also (buried) in the full listing of pairwise mean differences in the MIXED output. Note: this only makes sense if you have an interaction. With so many differences, the Bonferonni adjustment (Bon) method will be very conservative. In other words, it is probably getting extremely difficult to find any significant differences, even with a highly significant global test result. I highly recommend that you use a different adjustment. Try adjust=tukey; or adjust=simulate; . The latter is my favorite, but there are advocates of different adjusment methods. If you have a global interaction, you might want to get slices (simple main effects) for each region. That is, you can easily get MIXED to tell you if treatment has an effect for each region in a nice table. The lsmeans statement would then become: lsmeans region*condition/ diff adjust=simulate slice=region; I also think that your repeated statement is not right. Based on what you wrote, both condition and region are repeated measures. You are getting a simple variance-component model, and at a minimum, you need to account for variability between and within subjects. One simple approach is to use type=CS as an additional option on your repeated statement. But, you probably want to explore more complex structures. For instance, there is a spatial orientation to region, so you could model this in the repeated statement, and there are several choices. If you pursued this later idea, there are numerous issues to consider, and it would require that you read about the subject of spatial repeated measures. You could start with the the excelellent book by Littell et al. (SAS for Mixed Models, 2nd edition). The book by Brown and Prescott on mixed models in medicine is also excellent (and SAS based). ... View more

The intro to the LOGISTIC procedure (User's Guide), and example 51.4 (v9.2) deal with generalized logits (not ordinal) and the multinomial problem. But I am guessing that you have a different modeling approach in mind. In that case, there is no variable selection method. But as I wrote above, I would be cautious of any automatic variable-selection method. ... View more