You get ddf=1 with the KR method when things blow up with the variance-covariance parameters. When the complex calculations for the KR method give a nonsensical value, including a value less than 1, MIXED and GLIMMIX both give 1 for the denominator df. Ill-conditioned matrices are common culprits. I bet you have one or more warnings in your Log. I am thinking your model is overparameterized in terms of variance-covariance parameters. I know that arma(1,1) is especially difficult to fit. And as indicated by Steve, R-side covariance terms with a negative binomial can lead to all kinds of challenges. ... View more

Be careful. The default estimation method in MIXED is REML. That means, you can only compare AIC statistics for models with different random effects and the same fixed effects (the part in the model statement). If you want to compare models with different fixed effects, you will have to use method=ml in the procedure statement. Results could be a bit biased with small sample sizes. The null model is easy, nothing before the slash. model N = / solution ddfm = kr; ... View more

I can't do it in simple terms.In general, mixed models require an iterative approach to find the parameters that give the maximum value of the log likelihood (or minimum value of -2 times the log likelihood). Most optimization approaches use either the first or second derivative of the log likelihood with respect to the parameters in the iterative process. The Nelder-Mead simplex method (NMSIMP), however, does not use derivatives. It may take many-many iterations to converge, and the whole process can be very slow. It may become impossibly slow with many data points. But it is one of the valid methods. The second derivatives are calculated at the end of the optimization in order to estimate variances. See the Shared Concepts and Topics chapter in the SAS/STAT User's Guide (under NLOPTIONS) to learn more. ... View more

You should try HPREG procedure. This is designed specifically for high dimensional fixed-effects modeling. It is only found in the newer releases of sas. ... View more

I am guessing you have unbalanced data, with a lot of cells with no observations (in the three-way table). I usually see this problem with the default ddf method (ddfm=containment).  You could use ddfm=kr (for Kenward-Roger method). This will work for the default estimation method. However, the default estimation method is probably not the best choice for binary data because of bias of parameter estimates. I would suggest you use the option method=laplace on the procedure statement. But if you do that, you cannot use ddfm=kr.  If you felt you had enough observations, you could use chi-squared wald statistics instead of F statistics for hypothesis testing and confidence intervals. Just put CHISQ as an option in the model statement. ... View more

Then my original comments are appropriate. ... View more

Actually, since you are using the beta distribution, you are treating you data as continuous, but bound by 0 and 1. (I missed this before). So, the idea of overdispersion is not relevant. That is, the beta has a scale parameter that is estimated, so there is no real relevance to the chi-square/df statistic in you case (the way I am looking at it now). ... View more

One point: the goal is not to achieve a 0 for chi-square/df. Lack of overdispersion is indicated by chi-square/df = 1. Values less than 1 indicate underdispersion. It is OK to have a value less than 1, but you should not be including terms to specifically achieve this. For some model formulations (some choices for the random effects structure), the model fit will guarantee a value of 1. The NMSIMP method does not use derivatives in the optimization. It if fine to use, but is usually very slow to converge. ... View more

Some would argue against performing all possible pairwise comparisons, because too many of these are not of direct interest. However, I know the value of these multiple comparisons in many disciplines. The Bonferroni will be way too conservative with 10 treatments. I think simulation-based adjustments are best. Use lsmeans treatment / pdiff adj=simulate; ... View more

Because your response variable is a variance, I recommend that you model it as a gamma distribution, with a log link. That is, use GENMOD or GLIMMIX, and choose dist=gamma and link=log. The gamma often works very well as an approximation for the true distribution of a variance at small (finite) sample sizes. You can still get predictions for the original scale in an output file. For instance, proc glimmix ; model var = .... / dist=gamma link=log s ; output out=pred pred(blup ilink)=predicted; run; There are many other options for the output file. Schabenberger and Pierce (2002 textbook) give an example of a regression analysis of a variance dependent variable using this idea (but with GENMOD). ... View more

Here is the mixed code for one year (split plot with two whole plots). proc mixed; classes block a b c; model y1=  a|b|c ; random block block(a b); lsmeans a|b|c / alpha=0.01 diff adjust=simulate ; run; Random effects are only listed in a random statement. All the tests are conducted correctly and automatically (no TEST statements, etc.). See how easy it is. Duncan adjustments are not available, and are not recommended. I put in simulation adjustment, which is now considered a good choice. I actually think you should use a Kenward-Roger df adjustment also. Change model statement to: model y1=  a|b|c / ddfm=kr ; Now, we can consider year to be a repeated measure within the block*a*b*c experimental units (you said the same plot is observed for each year). I consider year to be a fixed effect. With only two years, the covariance structure has limited structure, so we can consider compound symmetry. This is the same then as a split-split plot with blocks. I doubt if you want all the interactions with year (you will never be able to interpret all the 2x,3x, and the 4x interaction). But I will write the model (one possibility) with the four way interaction: proc mixed; classes block a b c year; model y1=  a|b|c|year / ddfm=kr ; random block block(a b) block(a b c); lsmeans a|b|c / alpha=0.01 diff adjust=simulate ; *you can add more LSMEANS statements here, say for specific year interactions; run; The above assumes that the data are stacked, with a year variable added and a separate record for each year. If you had more than two times, you must consider more complex covariance structures. A more general approach for any number of times for the repeated measure is to use: proc mixed; classes block a b c year; model y1=  a|b|c|year / ddfm=kr ; random block block(a b) ; repeated year / subject=block*a*b*c  type=ar(1) ; *there are many choices for type; lsmeans a|b|c / alpha=0.01 diff adjust=simulate; *you can add more LSMEANS statements here, say for specific year interactions; run; or proc mixed; classes block a b c year; model y1=  a|b|c|year / ddfm=kr ; random block block(a b) random block(a b c) ; repeated year / subject=block*a*b*c  type=ar(1) ; lsmeans a|b|c / alpha=0.01 diff adjust=simulate; *you can add more LSMEANS statements here, say for specific year interactions; run; I expect that some variance components will be 0. One approach is to then remove them. ... View more

You need to better explain your design. Do you have a, b, and c, plus year, or is year one of the letters in your code? A split plot would normally have two fixed effect factors, not three. It looks like you have two factors crossed at the higher level (a and b), and then c is the sub-plot (within the axb combination). Is this right? ... View more