For problems like this, it may be helpful to start with a simpler model, such as one with no random effects. This will tell you if you have some problems with the coding of the data (for the fixed-effects part) , or if you have evidence of finite mixtures. Look at the residual plots. Then start adding the relevant random effects. You do have problems with your random effect syntax. First of all, with negative binomial, you automatically have an overall residual (R-side) scale parameter for the conditional distribution. But then in your original post you have two separate random statements that are both coding the R-side residual scaling parameter (either using the RESIDUAL option or the _RESIDUAL_ keyword). This is creating all kinds of conflicts and overparameterizations. Plus, I think you will get very strange results (and maybe not interpretable results) by trying a temporal autocorrelation structure for the R-side residual. If you read the recent discussion in the SAS/STAT discussion board, you will see some arguments against R-side repeated measures for GLMMs (open to debate). Also, the chol option won't do anything here because you have a simple variance component term for this statement. Start with a simpler random structure: random int bird / sub=nest; which is equivalent to random nest bird(nest); At least, this will let you get started. If this works,then you can consider more complex models (if needed).
... View more
