For valid statistical inference, it is important to select an appropriate statistical model. In the analysis of capture–recapture data under the closed-population models of Otis et al. (1978), information theoretic and hypothesis testing approaches to model selection are not practical, because some of the models have likelihoods with nonidentifiable parameters. A further problem is that, for some of the Otis et al. models, multiple estimators exist but there is no objective basis for deciding which estimator to use for a particular dataset. In CAPTURE, a computer program for estimating parameters under the closed models of Otis et al., a linear discriminant classifier is used to select an appropriate model. This classifier frequently selects the incorrect generating model in simulation studies, and it provides no guidance on which estimator to use once a model has been selected. In this study, we develop new classifiers for selecting the best estimator (as opposed to the generating model) and evaluate their performance. In addition, we investigate an estimator averaging approach to estimation that is a modification of the model averaging approach described by Buckland et al. (1997). We found that, in general, the overall performance of the new classifiers was unimpressive. In contrast, the estimator averaging approach we investigated performed well.

Key Words
Classification; Estimator averaging; Linear discriminant analysis; Model averaging; Model selection; Multinomial logistic regression; Population estimation.

Thomas R. Stanley is Wildlife Research Biologist, U.S.G.S Biological Resources Division, Midcontinent Ecological Science Center, 4512 McMurry Avenue, Fort Collins, CO 80525. Kenneth P. Burnham is Statistician, Colorado Cooperative Fish and Wildlife Research Unit, Colorado State University, Fort Collins, CO 80523.