this talk is about model selection for clustered data. a conditional akaike information, cai, has been defined for linear mixed models when the inference is on cluster, rather than population, parameters. we extend this definition to generalized linear mixed models and derive an asymptotically unbiased estimator of cai. this estimator is applied to a cancer data set. alternatively, cai can be estimated using a nonparametric bootstrap. we compare the two estimators in simulation