Hello All, Sorry for the long discussion title. I've found a couple previous discussions on this warning, but none of the solutions proposed seem to fix my problem. So, briefly here is a description of the data I'm trying to analyze, the problem I'm having, and the solutions I've already tried: Data is from a pollination experiment focusing on 2 species. I had - 10 plots randomly assigned to 2 main treatments: Unmanipulated, and Removal - within each plot, 10 randomly selected individuals were monitored - of the 10 individuals in each plot, 3 were assigned to a pollen supplementation treatment, the other 7 were left unmanipulated. - I collected seed count data for every flower on every individual as my response variable, and recorded flower position on the flowering stem as an important covariate. I am treating individuals as subjects for which i have repeated measures (multiple flowers per individual). I have strongly correlated within-individual responses with respect to flower position, so, I have tried to account for those correlated responses with my R-side covariance structure as follows: proc glimmix; class treatment plot poll_supp flowpos_1; model nseeds = treatment poll_supp flowpos treatment*poll_supp treatment*flowpos poll_supp*flowpos treatment*flowpos*poll_supp / dist=negbin offset=lntovules ddfm=kenwardroger solution; random flowpos_1/ subject=ind type=ar(1) residual; random plot; run; However, no matter how I tweak this code (e.g. dropping the higher order interactions, using different dist=, ddfm=, with or w/o offset=), so long as I include "random flowpos_1/ subject=ind type=ar(1) residual;" it blows up and gives me the Warning: obtaining minimum var. quadratic unbiased estimates as starting values for the covariance parameters failed. At first, I thought it was just a d.f. deficiency, so I dropped interactions, and used a copy variable flowpos_1 to order the observations for the random statement, using flowpos as a continuous variable instead (this is actually a preferable formulation, but I originally tried it to sneak back some d.f.'s). Plus, the dataset has ~407 observations, and fingermath on the number of parameters being estimated checked out ok, so I think I am ok for d.f. My second thought was that this is just a numerical issue because of strong imbalance in the observations at different levels of flowpos (many individuals with 1-10 flowers, very few with 11,12; actually only 2 observations at 12, w/ variance=0)... so I tried subsetting the dataset and dropping the observations at the highest levels of flowpos... but am still having the same issue... So... My conclusion is that I am writing the code for that random statement incorrectly, but can't find any good examples to follow. If anybody has any advice or suggestions I would appreciate any help I can get. I know there are alternative formulations to account for my within individual correlated responses, but this seems like the formulation that does the best job of staying true to the biology in the experiment, and if I can avoid dropping the R-side covariance structure to account for within-individual correlated responses with respect to flowpos, I'd like to. thanks again for any help or suggestions. cheers, Colin
... View more