This is a guess only, but if both samples have the same number of clusters (and the data is simply duplicated), I would expect the means (proportions) and standard errors to be identical.   SteveDenham ... View more

From the PROC GLM documentation:   Note: PROC GLM uses only the information pertaining to expected mean squares when you specify the TEST option in the RANDOM statement and, even then, only in the extra F tests produced by the RANDOM statement. Other features in the GLM procedure—including the results of the LSMEANS and ESTIMATE statements—assume that all effects are fixed, so that all tests and estimability checks for these statements are based on a fixed-effects model, even when you use a RANDOM statement. Therefore, you should use the MIXED procedure to compute tests involving these features that take the random effects into account; see the section PROC GLM versus PROC MIXED for Random-Effects Analysis and Chapter 81: The MIXED Procedure, for more information.   This applies to interval estimates as well, so that GLM doesn't produce CI's as you might think of them.  I thought you might be able to add the CLPARMS and SOLUTION options to the MODEL statement and get something, but that appears to give CI's for each level of the effects.   I guess this all boils down to "Don't use PROC GLM for this particular task.  Use PROC MIXED."   SteveDenham ... View more

This code fits two variance components:   random int episode/subject=ID and is equivalent to:   random ID episode*ID; While this code fits one variance component    random int /subject=episode(ID) and is equivalent to   random episode*ID; The key here is that the matrix representation of nested effects is identical to the matrix representation of crossed effects (see the documentation of the RANDOM statement).   It comes down to partitioning the variability between ID and episode*ID, with the remainder, if any, in the residual.  All of this is going to be very data dependent.   SteveDenham         ... View more

If you have 3 components, you might try doing 3 stepwise Satterthwaite calculations (A with B, and then the result with C; A with C, and then the result with B; and B with C, and then the result with A), and then averaging those to get approximate df.  Sounds a bit brute force, though.   SteveDenham ... View more

This one isn't behind a paywall, and looks like it covers much of the same material as the other Francq paper.   SteveDenham ... View more

Not real sure how this is going to work.  It appears to me that you could get Var(interceptR) by adding solution as an option for the random statement.  From the BLUP, you could then calculate the combined variance.   SteveDenham (this is a real shot in the dark, though) ... View more

You have three responses for resp_never_married - missing, 0 and 1, so the program assumes that the variable is not binary.  Try removing the missing response records and see it that improves things.   SteveDenham ... View more

It generally is a good idea to account for site variability in an analysis, but be cautious.  If there are a small number of sites, you may not be able to estimate a variance component due to site.  You might need to fit it as a fixed effect.  This is a fairly common occurrence in ecological modeling.   SteveDenham ... View more

I believe you have identified the source of ill-conditioning, and the log transform is often used with income levels, as the data is likely long-tailed to the right.  You could fit a lognormal distribution using PROC GLIMMIX, but the results should be nearly the same.   SteveDenham ... View more

It appears to me that there is no repeated effect in your design - there is one outcome for each of the 4950 pairs.  That leads to another question - how is exposure measured?  The only estimation available is the marginal effect of exposure on outcome, averaged over all the pairs (I think).  What do you get if you remove the REPEATED statement?   SteveDenham ... View more

Why are you categorizing a perfectly good continuous variable?  From a statistical point of view, it is the equivalent of taking a Lamborghini into the shop and asking them to cut it up into pieces that might fit into the garage better.  See this link:https://discourse.datamethods.org/t/categorizing-continuous-variables/3402    SteveDenham ... View more

Your design implies that you have 4950 pairs that are correlated (100 take 2).  How many total observations do you have?  If it isn't at least 49,500, GEE might not be the best tool.  In fact, you may need to analyze your data in some other fashion designed for multinomial responses (FREQ, LOGISTIC, GENMOD, CATMOD) that would use the levels of pair and the levels of exposure (and their interaction, if possible) as factors.   SteveDenham ... View more