This is the difference seen when comparing an "intent to treat (ITT)" analysis to a "per protocol (PP)" .  The first (and your first) uses all of the data to create the matrices needed to solve the mixed model equations, while the second (and your second) uses only those subjects for which the response variable is known.  In a repeated measures analysis, the PP method implies that records with a measured response, but a missing value for any of the covariates (fixed or random) will be included in the analysis.    SteveDenham ... View more

If you want marginal estimates, try PROC GENMOD.  If you want conditional estimates, try PROC GLIMMIX.  If you know the structural equations/path model, try PROC CALIS.  All of these handle repeated measures of a binary response.  Selecting which to use depends very much on what @ballardw  and @PaigeMiller  are asking for - a concise, but complete, statement of the research question.   SteveDenham ... View more

The second example in the LOGISTIC Procedure documentation has something that may be what you want, using the ODDSRATIO and CONTRAST statements.  Here is the online link:  https://documentation.sas.com/doc/en/statug/15.2/statug_logistic_examples02.htm    SteveDenham ... View more

Hmm. That sounds a bit like PROC TRANSREG, but much of the model "selection" will have to be done manually based on output. However, you can get optimal transformations of the variables which may be what you are looking for.   SteveDenham ... View more

To get what you want, you will also need some estimate of the time point variances and the covariance between time points.  At that level, you may be happier using simulation to determine sample size, or using the method outlined by Stroup in this paper   https://citeseerx.ist.psu.edu/viewdoc/download;jsessionid=1DD150BA59B7EB27C5B890D4154BC3FE?doi=10.1.1.546.1468&rep=rep1&type=pdf    The other thing I would offer is that your pilot data has most of the VAS scores near the middle of the range.  At that point, you can consider the residuals to be approximately normal rather than using a very difficult to estimate ordinal multinomial (missing values will almost certainly be an issue - I doubt that any of the timepoints has values at all 101 points from 0 to 100.   SteveDenham ... View more

In this case, you can probably get estimates if you specify 'site' in the CLASS statement.  The more I look at this, I really need an explanation of enroll_order and how it is coded. I don't think your RANDOM statement correctly models the design in your data, as I don't see how to include 'enroll_order' as a subject for each site.  Without more information, I suspect that there is a different enroll_order for each site, which would mean that the two variables are confounded.  Please include more information regarding the data structure, and also include any messages in the log that may be generated.   SteveDenham ... View more

For your GLIMMIX code, I would add enroll_order to the CLASS statement. I am somewhat baffled by the GENMOD code, though.  The CLASS statement there looks to me as if it treats your response variable (score) as a multinomial.  Given that, I am not sure that the data follow a Poisson distribution, but rather follow an ordinal multinomial distribution.   SteveDenham  ... View more

Given all that has been presented by @StatsMan and @Rick_SAS, I believe that you can model Left Eye and Right Eye as repeated measures on the individual. In this case 'eye' is a fixed effect just as 'visit' is, but you have a chance to capture within subject variation for both of these. Here is the code you propose:   proc glimmix data=analysis plots=studentpanel; class sex usubjid eye visit ; model aval= visit|eye age sex / ddfm=kr s; random intercept / subject=usubjid ; random eye/ residual subject=usubjid (visit ) type=cs; lsmeans eye*visit / alpha=0.05 cl ; /*comapre visit to their baseline */ LSMeans eye*visit/ slicediff=eye slicedifftype=control( 'Left' 'Baseline') alpha=0.05 cl ; /*compare right eye and left eye at diffrent visits */ LSMeans eye*visit/ slicediff=visit slicedifftype=control( 'Left' 'Baseline') alpha=0.05 cl ; run; Adding one more RANDOM residual statement would accomplish what I think you want:   proc glimmix data=analysisplots=studentpanel; class sex usubjid eye visit ; model aval= visit|eye age sex / ddfm=kr s; random intercept / subject=usubjid ; RANDOM VISIT/RESIDUAL SUBJECT=USUBJID TYPE=AR(1); random eye/ residual subject=usubjid (visit ) type=cs; lsmeans eye*visit / alpha=0.05 cl ; /*comapre visit to their baseline */ LSMeans eye*visit/ slicediff=eye slicedifftype=control( 'Left' 'Baseline') alpha=0.05 cl ; /*compare right eye and left eye at diffrent visits */ LSMeans eye*visit/ slicediff=visit slicedifftype=control( 'Left' 'Baseline') alpha=0.05 cl ; run; I will say this though. We have had more success treating eye as a fixed effect (as pointed out by @StatsMan and @Rick_SAS ), which avoids a LOT of failure to converge and G matrix non-positive definite issues. Something like this:   proc glimmix data=ansdata_&n. ; nloptions maxiter = 5000 ; class &grp_no. &time. &animal. eyen; model &value. = &grp_no.|&time.|eyen / ddfm=BW; random &time. / type = &cov. subject=eyen(&animal.); %if "&cov." = "SP(POW)(T)" %then %do; random intercept / subject=eyen(&animal.); %end; Note that the macro variable &grp_no. may include sex as in your code, but that there is an opportunity to include interactions that may not be apparent in your model.   SteveDenham       ... View more

Two things you might try.  First, lmer uses maximum likelihood as a default, while MIXED uses restricted maximum likelihood (REML).  So you might add method=ml to the PROC MIXED statement.  The other is to change the random effect to specifically fit random intercepts:    random intercept/subject=id That is as much as I really know about the two platforms.   I do have a couple of questions that could affect your analysis.  Is there any reason to suspect autocorrelation of the residuals?  If so, a type= option to model that might be in order.  The second has to do with specification of repeated effects in lmer.  It looks to me as if you are modeling week as a random effect, rather than modeling correlation in the residuals.  The latter is what the REPEATED statement in PROC MIXED does.  I could certainly be wrong on this, but if you want to check this, then change the REPEATED statement to a RANDOM statement, and see if the results are closer to the lmer results.   SteveDenham   ... View more

If you are going to use the ddf option on a split-plot design, use the df from the Type3 analysis.  I would then apply the empirical option.   SteveDenham ... View more

You might want to try a conditional model using PROC GLIMMIX, but if all the missingness is in the dependent variable, that probably will not help much. With responses that are MAR (missing at random), your choices come down to fitting the complete case data only, or estimating the response level in the missing response cases using the OUTPUT statement to get predicted values for the missing responses, and replacing the missing values with the most likely level and then refitting the model.  You could also look into using PROC SCORE to estimate the missing responses, but that looks pretty rocky to me, given that your response is an ordinal level.   SteveDenham ... View more