This will not work:   lrecl=30000000000   "In SAS 9.4, the default value for LRECL= is 32,767.... The maximum length is 1G."   https://documentation.sas.com/?docsetId=hostunx&docsetTarget=p18r7xw5gg6is3n127xdj69d6b4p.htm&docsetVersion=9.4&locale=en ... View more

You ought to have a look at box plots and diagnostic graphs.   "...Levene’s test or any other test for homogeneity of variance should not be used as a diagnostic for the assumption of equal group variances that underlies the usual analysis of variance..."   https://support.sas.com/documentation/cdl/en/statug/63347/HTML/default/viewer.htm#statug_glm_sect056.htm ... View more

NPV and PPV are sensitive to prevalence, whereas likelihood ratios are not. Instead of dramatically reducing your prevalence from 75% to 6%, why not calculate the likelihood ratios as demonstrated on this page:   http://support.sas.com/kb/24/170.html ... View more

Code to do this is available here:   https://documentation.sas.com/?docsetId=grstatproc&docsetTarget=p0relmtk2s8ac9n13bnmicsgsmwu.htm&docsetVersion=9.4&locale=en ... View more

Better post it at Stat forum and calling @StatDave  ... View more

   Aska, i think you can use this code for binomial distribution with repeated measures: proc glimmix data=A; class subjectID treatment day; model reactivity=treatment day treatment*day / dist=binomial DDFM=KR; random subjectID; lsmeans treatment day treatment*day/ pdiff lines; /* if there is a significant interaction then you can compare the treatment effects for each day*/; lsmeans treatment*day/sliceby=day pdiff lines; run; Thank you, Bikash ... View more

Since your response is binary (positive vs negative), PROC MIXED is definitely not what you want to use since it expects a normal response. Cochran's Q tests that the marginal probabilities (probability of positive in each sample) are equal, but since you seem to really be after assessing conditional probabilities (such as probability of positive in sample 4 given result in sample3), I don't think it is what you want either. One thing you could consider are transition models as described in this note.  ... View more