Well, REF=LAST is the default, so the analysis you propose should set the estimate for ISCED_gr=3 to zero, which establishes it as the reference category.   SteveDenham ... View more

I recommend getting a copy of SAS for Mixed Models. Either the 2nd edition or the newer Introduction and Basic Applications edition. Both have chapters on power and sample size determination (the newer version has two chapters on this topic). In any case, you feed in the variance components and an exemplary data set that builds in the differences you want to be able to detect, and run PROC MIXED holding random effects fixed with a PARMS statement. You then post-process the results to get an approximate noncentrality parameter. With that in hand, you can calculate the power. There are several worked examples.   SteveDenham   ... View more

I see that you mentioned PROC REG, and are wondering why PROC MIXED doesn't produce the parameters you are used to. Well, it does. In the fixed effect solution table, the Estimate column gives almost what you are expecting. for continuous variables, the estimate is just like what you see from PROC REG. However, the use of categorical variables through the CLASS statement leads to a '0' estimate in the Solution table, which is pretty confusing at first. The reason is that the parameterization in PROC MIXED is non-full rank, i.e. one level of a class variable is set to zero. The other levels of that variable are presented as the difference from the level set to zero, called a reference level in some cases. The documentation for SAS/STAT has some great introductory material that goes through this, so look at Introduction to -- Statistical Modeling with SAS/STAT Software; Regression Procedures; Analysis of Variance Procedures; and Mixed Modeling Procedures, and The Four Types of Estimable Functions. I think all of your questions have answers there.   SteveDenham ... View more

From the SAS/QC 15.2 documentation, it appears that you need to add SSYMBOL=SBAR as an option to the XSCHART statement (although SSYMBOL=SBAR is called out as the default). If  the central line is not constant, only the symbol is displayed. That sort of looks like what you currently have being printed, but your central line looks constant so that shouldn't be an issue.   Subgroups? The Turbines example in the PROC SHEWART documentation looks to me like the subgroup is equal to the full sample size (n=20). If you have multiple subgroups then maybe that is the source of why the 'value' part of the Sbar isn't getting printed.   SteveDenham ... View more

Have you considered using PROC GLIMMIX, as in the following:   proc glimmix data=MAST_long_VAS; class PPID Time; model BC_stress = Time Time*Time / solution; random _residual_ / subject=PPID type=ar(1); run; In large part, this is adapted from the example - Example 22.10 Correlation and Covariance Matrices in Using the Output Delivery System documentation.   SteveDenham   ... View more

I feel like I had pretty good luck with the following code:   proc glimmix data=tab1 noprofile; class gr experiment; model var1 = gr experiment/ddfm=contain; random experiment/residual type=ar(1) subject=gr; lsmeans gr experiment/diff=control; run; You cannot fit this as a full factorial model as it is completely saturated.   Another approach would be to fit a GEE model. I used PROC GEE (SAS/STAT 15.2) to have access to the Type3 option for testing the model effects.   proc gee data=tab1; class gr experiment; model var1 = gr experiment/type3 ; repeated subject=gr /type=un; lsmeans gr experiment/diff=control; run; The results are different, but reflect the different approaches of the two procedures. In GLIMMIX, the lsmeans and tests are conditional (subject specific) while for GEE they are marginal (population averaged). Both are appropriate, but the choice depends on your research question.   SteveDenham  ... View more

By any chance, is cleartitle a level within Mktval2? If that is the case, it may be getting used as a reference level. If not, try adding the type3 option to your MODEL statement and PARAM=GLM to your CLASS statement.   But as @ballardw  points out, it is very difficult to answer a question where we don't know exactly what you are doing currently, and do not have access to the output.   SteveDenham ... View more

For GLIMMIX and HPGENSELECT, you must specify the output variable in the CLASS statement if you are fitting a generalized logit link to a multivariate distribution. For LOGISTIC, it is as @StatDave says - don't put the output variable in the CLASS statement. I don't know why it isn't consistent.   SteveDenham ... View more

The only way I know of specifying the reference level for the response variable is to shift to fitting a generalized logit to the multinomial distribution. If you go that way, you can specify any particular level of the response variable as the reference using the ref='  ' method. The source of the NOTE: regarding the reference category in the response variable is due to the link chosen. As I mentioned in the first sentence, you have to specify LINK=GLOGIT for it to apply the reference level. There are other PROCs that operate similarly (HPGENSELECT, GLIMMIX for example).   Just thought of another way, but it requires formatting the levels of the response variable. Just set up your format so that the level you want as the reference is either LAST (default) or FIRST (needs a REF=FIRST in either the MODEL or CLASS statement).   SteveDenham   ... View more

To answer the question of the validity of collapsing categories, it is critical that we know what the research question is, and how you expect the data you have to be used in answering the research question.   The reason I pose it that way is that the only answer I can offer given the information at hand is "I don't know enough about what you are doing to come up with an appropriate answer." For some research questions, it would make sense to collapse the AB testing to a binomial; for other questions, you may need to stay with the multinomial.   SteveDenham ... View more

Given that, you should mark @Ksharp 's post where he suggests a SLICE statement as a correct answer.   SteveDenham ... View more

Not necessarily incorrect. For instance, a Poisson distributed variable might be detected by looking at the mean and variance of the variable. If the two are equal, or close to equal, a Poisson might be a good starting point. Then you could examine the residuals from a model with a log link by various plots, such as QQ to see if you get a straight or almost straight line.   SteveDenham ... View more

The only way I can imagine fitting a mixed model, rather than a fixed effects model, that includes the predictors you have is if the census regions included are not exhaustive of all possible census regions to which you wish your inferences to apply (G-side effect), or that some sort of spatial correlation is to be applied to that variable (R side effect). In either case, a generalized estimating equation (GEE) where census_region is the clustering variable may be more tenable in terms of model convergence and run times.    SteveDenham ... View more

I still don't see how this is going to give you a "pooled" p value. It should give you 15 adjusted p values. I might have misinterpreted the request but I thought you were looking for a single pooled p value.   SteveDenham ... View more