You can explore the Kronecker product structures for TYPE= on the REPEATED statement. That is a handy construct for two-dimension repeated measures. You can set up an unstructured covariance between the eyes of a patient and CS, AR(1), or (on occasion) UN for observations within an eye. The model you have now is used quite often as well. You could try adding covariance within an eye as an extra random effect, with random int eye / subject=ptid; That RANDOM statement correlates observations from the same patient and further correlates observations from the same eye within a patient. You could fit AR(1) as a structure on the REPEATED statement you have now with this RANDOM statement.  ... View more

If you can post the code you are using, with and without the interactions, and the output from both then that would give us a better understanding of what is happening here. ... View more

Any success with your overnight run? Were you able to try HPMIXED? ... View more

Posting your results and providing more detail about your data (replications, specifically) will help us diagnose the problem ... View more

You may need to specialize some error terms here. This example in the GLM doc will get you started. PROC MIXED is a good alternative to GLM in these situations, too. ... View more

Once the model contains random effects that are non-nested (as you have with MOVIE and ID), the model becomes much harder to estimate computationally. If all random effects are nested, then MIXED can process through the V matrix in blocks. If the random effects are not nested, it processes the entire V matrix at once.  You can try PROC HPMIXED and see if that improves performance. You can also try starting your model in MIXED late in the day and let it run overnight, if you have not done so already.  ... View more

There is a lot going on in this data. Your response, INTENSITY, appears to be measured on a 7 point scale. If that is correct, you need to decide if you want to approach this response as multinomial or continuous. Your modeling choices are a little trickier with a multinomial response (PROC GLIMMIX can handle this) than with a continuous response. While some would argue that with an integer response that you should take a multinomial approach, others would be ok accepting a 7 point scale as continuous.  With the continuous approach, you can look at MIXED or GLIMMIX as your modeling procedure.  The next choice will be how to approach the predictors in your data. You have CLIP within MOVIE (6) within TYPE (2). It is fairly certain that you want to treat TYPE as a fixed effect. You also have GENDER as fixed effect.  If you want to generalize your results to a larger set of movies, then treat MOVIE as a random effect. With MOVIE as a random effect, CLIP within MOVIE will also be random. A statement like random int clip / subject=movie; will handle both of those random effects in MIXED.  You also have responses that come from the same rater id. With movies changing across the rater id's, you do not have another levels of nesting here. This makes the model a bit more memory and time intensive to estimate.  A REPEATED statement like repeated / subject=id type=cs; is your safest, easiest approach. Since the raters do not see the same clips, fitting a CS structure would appear to make the most sense without knowing anything more about the data.  Your CLASS statement will contain TYPE, ID, CLIP, MOVIE and GENDER. Use TYPE and GENDER as fixed effects on the MODEL statement.  Hope that helps you get started! ... View more

The covariance structure may be too complicated for binary data. If you are having trouble with convergence, then drop the R-side fit and see if that helps. Also switching the optimization to NRRIDG can help as well. Using LAPLACE or QUADRATURE gives you models you can compare, using the fit statistics.  ... View more

Estimate statements for slopes are tricky. It gets even trickier when you have many interactions. Harder still is averaging slopes over one of the model CLASS effects. The best strategies for getting to your end result is to build up your ESTIMATE statements using more simple tests and using the /E option on the ESTIMATE statement to make sure you have coefficients in the correct place. It is scarily easy to write an ESTIMATE statement that is giving you an answer but that is testing the wrong hypothesis! First, try to write ESTIMATE statements for each of the 6 slopes on genotype for the 2 sites. Those ESTIMATE statements will involve the main effect of CONCENTRATION and the interactions of SITE*CONCENTRATION, GENOTYPE*CONCENTRATION, and SITE*GENOTYPE*CONCENTRATION. The coefficients on the ESTIMATE statement will all be 1's here, in the appropriate spots for each of these 4 effects. Again, that /E option will show you if you have your coefficients in the right spot.  Then, you can average the coefficients over the genotypes in a site to get your final result. With 6 genotypes, you can either use .1666666 as the coefficients on the G*C and S*G*C interactions, or use 1 and the DIVISOR=6 option on the ESTIMATE statement. The coefficient on C and S*C would be 6 in this case using that DIVISOR= option. If you get a non-estimable result for that ESTIMATE, check the /E option output to make sure you have the coefficients in the right place. Check that output even if you don't get a non-estimable result to make sure you are testing the correct hypothesis.  It takes a lot of experience to get estimates like this correct. Working with an experienced statistician is always a good idea when writing post-hoc tests like these.   ... View more

You almost had the nesting syntax. For that second level of nesting, you need a(b c) instead of a(b(c)).  ... View more

With TYPE=CS on the 2nd RANDOM statement, you are likely getting a message in the log that says either the G matrix is non-positive definite or the Hessian is non-positive definite. You are also likely to see that the CS parameter in the GLIMMIX output is 0, or that if you remove TYPE=CS then you will see that two of the parameters in the model with TYPE=CS sum to one of the parameters in the model without TYPE=CS.    In the slide you reference, the second RANDOM statement with SUBJECT=PTID correlates all of the observations from the same patient. The first RANDOM statement, with SUBJECT=EYE(PTID) and leaving off TYPE=CS, correlates the observations from the same eye within a subject. If you use the V option as I suggested earlier, you will see a block diagonal covariance structure that shows this correlation pattern. Here is an example:   This is the result of the V option for 2 subjects. The first 3 rows and columns are for 3 observations on eye 1 for subject 1. Rows and columns 4-6 are for 3 observations on eye 2 for subject 1. This structure was obtained without using TYPE=CS on the RANDOM statement for SUBJECT=EYE(PTID). You get a correlation within the eyes and within all observations on the same subject.  ... View more

Your model would be reasonable with multiple follow-up visits. With a single non-baseline response, you will have difficulty estimating the random effect for subject_ID(clinic). If you get a message that the g-matrix is non-positive definite, then you should consider dropping that RANDOM statement.    For a good write-up on approaches to this kind of data, see https://www.lexjansen.com/phuse/2017/as/AS07_ppt.pdf.   The cLDA model works well when you have multiple follow-up visits. I am not sure of its effectiveness with a single follow-up. ... View more

When you fit random effects on the G side (which you are doing with your RANDOM statement in GLIMMIX) then specifying something like random intercept / subject=id; will fit a common covariance to all the observations with with same id value. This fits a block-diagonal structure through Z'GZ. You do not need a covariance structure on the RANDOM statement to achieve this.   When you fit random effects on the R side (using the RESIDUAL keyword on the RANDOM STATEMENT in GLIMMIX or the REPEATED statement in MIXED) then you need TYPE=CS to get this block diagonal structure. V=Z'GZ + R, so you need to specify the covariance structure for the R part here.    You can confirm this behavior through the V option on the RANDOM statement. V shows blocks of the V matrix when GLIMMIX (or MIXED) can partition V into blocks (it needs the same subject or nested subjects on multiple RANDOM statements to do this). Use V=2 to see the block for the 2nd subject.  ... View more