For a model of this complexity, basically your options are simulation or a clever use of PARMS / HOLD that Walt Stroup has written about:   PROC GLIMMIX as a Teaching and Planning Tool for Experiment Design   Ch 16 in Generalized Linear Mixed Models: Modern Concepts, Methods and Applications   Even with this tool, you may still be a bit lost because of the challenge of specifying one (or many) alternative hypotheses in a four-way factorial treatment design. I would be, anyway.     Good luck!   ... View more

If you are assuming a normal distribution, then the OLS estimators are the same as the ML estimators. Consequently, GLM, GENMOD, and GLIMMIX will produce the same results. (GENMOD and GLIMMIX are able to accommodate non-normal distributions, but that is apparently not your situation.)   Some variant of the SLICE option or statement is available in all of these procedures, and the SLICEDIFF option is similar. You can learn a lot by trying the different Proc approaches. I encourage you to spend the time to learn more about the process; blindly following someone else's template is dangerous if you don't know what you are doing and why. Documentation is your friend in this activity, and you can always ask questions on this forum. ... View more

Like @PaigeMiller I am not completely sure what you want. But perhaps the SLiCE statement would work:   https://support.sas.com/documentation/cdl/en/statug/63962/HTML/default/viewer.htm#statug_genmod_sect030.htm   something like (which I have not tested😞     proc genmod data=have; class year group IV3; model DV = year group year*group IV3 / solution; lsmeans year*group; slice year*group / sliceby=year diff cl; run;   Note that I have added "group" to the MODEL statement to maintain the interaction hierarchy principle, see   https://www.quora.com/Why-do-we-have-the-hierarchical-principle-in-adding-interactions-to-a-model-What-is-the-significance-of-it   The "coefficients" would be produced by the SOLUTION option on the MODEL statement; but keep in mind that the coefficients depend on the parameterization of the model and are not unique. In my opinion, the LSMEANS are more intuitive (and do not depend on model parameterization).   The SLICE option is also available for the LSMEANS statement in PROC GLM. The documentation tells you everything you need to know 🙂   ... View more

Regarding the WARNING statement: The WARNING is suggesting that you specify the <repeated-effect> in the REPEATED statement; see the documentation REPEATED statement. It would be something like   repeated mth / subject=blk*fert*nfix*id type=sp(pow)(c-list);   where c-list is a numeric variable (which means it would not appear in the CLASS statement).   Regarding CONTRAST and ESTIMATE statements: This paper addresses your questions in detail. In particular, consider using the LSMESTIMATE statement rather than the CONTRAST or ESTIMATE statement; LSMESTIMATE makes writing contrasts and estimates easier and less error-prone.   You didn't ask, but I would think that if you want to analyze growth "rate", then you would need to apply an appropriate regression model for the response over time. But I know essentially nothing about your study design or the definitions of your variables, so this is just speculation on my part.   ... View more

Given that your dataset name is "Thesis.data", you will have written, or soon will be writing, a Methods section that describes your experiment. If you are willing to share, that would be useful and would contain the information we need.   This is what I am guessing about your experimental design:   Apparently you have subjects. Subjects is a random effects factor. Each subject has a single value of ICC, BMI, and IDD, which are fixed effects factors.   Each subject is "subdivided" into two sub-units; sub-units are nested within subject. Sub-units is a random effects factor. One sub-unit is assigned to one level of SOURCE. SOURCE is a fixed effects factor.   Each sub-unit is "subdivided" into three sub-sub-units; sub-sub-units are nested with sub-unit. Sub-sub-units is a random effects factor. One sub-sub-unit is assigned to one level of CPG. CPG is a fixed effects factor.   CPG is nested within SOURCE only if the three levels of CPG for SOURCE=1 are different than the three levels of CPG for SOURCE=2. If you have only 3 levels of CPG and these 3 levels are used with both SOURCE =1 and =2, then CPG is crossed with SOURCE. If these concepts of "nested" and "crossed" are unfamiliar to you, then you need to study the resources I suggested. Another good resource is Analysis of Messy Data Volume 1: Designed Experiments, Second Edition. Once you've done that, feel free to post code if you would still like feedback.     ... View more

I am not yet able to provide any advice about your model specification because your description does not clearly identify your experimental design. The resources that I suggested in my previous message will help you determine (1) your fixed effects factors, (2) your random effects factors, and (3) how levels of your fixed effects factors are (ideally) randomly assigned to levels of the appropriate random effects factors. Once these design elements are clear, we may be able to suggest appropriate model specifications. ... View more

When you are working with mixed models, it is important to clearly distinguish between experimental treatments (fixed effects factors) and experimental units (random effects factors).   In GLIMMIX, factors in the MODEL statement are fixed effects factors: they determine the mean of the response. Subjects in the RANDOM statements identify random effects factors that determine the variance of the response; we may use fixed effects factor names to identify levels of a random effects, which will be confusing until you fully understand how the GLIMMIX syntax works.   In your code, you are using SOURCE and CPG as both fixed and random effects factors, which is incorrect and suggests that your understanding of how to implement a mixed model is GLIMMIX is not yet sufficient. You'll find these references useful to study:   SAS® for Mixed Models, Second Edition   Generalized Linear Mixed Models: Modern Concepts, Methods and Applications (especially Ch 2, 7, and 😎         ... View more

Using simulation studies, Day and Quinn (1989) COMPARISONS OF TREATMENTS AFTER AN ANALYSIS OF VARIANCE IN ECOLOGY showed that the Tukey method is conservative with respect to Type I error rate; in Fig 1 you can see that the actual Type I error rate is about 0.02 when the nominal rate is 0.05 for their simulation scenario. Consequently, Type II error rate also is excessively increased (i.e., you lose more power). There's no perfect solution 🙂   ... View more

The INFLUENCE option produces internally and externally studentized marginal (not conditional) residuals.   ... View more