Hi, I posted here with a question about using PROC GLIMMIX on my own repeated-measures data, and got some great answers! Now I'm helping a colleague analyze his repeated-measures data, and I can't get it to converge. The response variable (response) is a count variable ranging from 0-19, but very left-skewed with many zeros and ones (to give you an idea: median=2, 95th percentile=8) Fixed effects: sex (M/F), treatment (4 levels, categorical), and trial. Trial is the proxy for time, and this is where things get complicated: subjects (=bird) were subjected to as many as 12 trials, but if they "passed", the trials would be stopped early. Some birds went through as few as 6 trials. Random effect: nest. Many of the birds came from the same nest (i.e., are related). Here's my code: proc glimmix data=MyData plots=(all) method=RSPL noreml IC=PQ plots=residualpanel order=data; Class bird sex treatment trial nest; tr = trial; model response = sex|treatment|trial / dist=poisson link=log ddfm=kr; random trial / residual subject=bird(nest) type=sp(pow)(tr); random int / subject=nest type=chol; random _residual_ / subject=bird(nest); nloptions maxiter=500 tech=congra; run; I have tried multiple variations of this code, including: - method=RMPL, MMPL, MSPL - dist = negbin, gamma, gaussian/log - ddfm = betwithin, contain, {default} - R-side covariance structures = ar(1), ante(1) - G-side covariance structures = un, {default} - removing bird from the class statement (the data is sorted by "bird trial") - I've also attempted running this model with all G-side random effects and using method=LAPLACE or QUAD. Nothing works. Either I can't get it to converge - in which case I get the "pseudo-likelihood update fails in outer iteration #" error message. Or it converges, but all the F values are higher than I would expect, and most-all LS means are "Non-est". I was also getting a "data label collision avoidance" error for a while, but was able to avoid that by entering "LABELMAX = 12400" to the ODS GRAPHICS statement. Any suggestions? Thanks in advance!
... View more