Appendix C. Supplementary details on the GLMM analyses.
Variation in area under the curve (AUC) across the 10 modeling techniques was also analyzed using a generalized linear mixed model (GLMM), with the AUC as the response. Technique was fitted as a fixed effect, and Species as a random effect. Analyses were performed using WinBUGS (Spiegelhalter et al. 2003a), which fits a Bayesian model. We assumed uninformative priors for all parameters, resulting in a GLMM that is equivalent to one fitted using standard maximum likelihood. Comparisons of the 10 techniques were summarized from 50000 Monte Carlo iterations after a burn-in period of 10000. The performance of Technique was summarized as the mean and standard deviation of the posterior distributions. The percentage of runs where the AUC for model A is greater than that for model B estimates the probability that the true difference between the models is greater than zero. This is a two-tailed test, and values close to 1 mean that model A's response is greater than that of model B, and vice-versa for values close to zero. The importance of the fixed effect Technique in the GLMM was assessed by change in the deviance information criterion (DIC, Spiegelhalter et al. 2003b) for the full GLMM compared with a second model where Technique was excluded from the GLMM. The DIC is the Bayesian equivalent of Akaike's information criterion, and rules of thumb suggest that changes in DIC of more than 10 units indicate that the excluded term had an important effect (Burnham and Anderson 2002, McCarthy and Masters 2005).
LITERATURE CITED
Burnham, K. P., and D. R. Anderson. 2002. Model selection and multi model inference: a practical information-theoretic approach. Springer-Verlag, New York, New York, USA.
McCarthy, M. A., and P. Masters. 2005. Profiting from prior information in Bayesian analyses of ecological data. Journal of Applied Ecology 42:1012โ1019.
Spiegelhalter, D., A. Thomas, N. Best, and D. Lunn. 2003a. WinBUGS user manual, Version 1ยท4. MRC Biostatistics Unit, Cambridge, UK.
Spiegelhalter, D. J., N. G. Best, B. P. Carlin, and A. van der Linde. 2003b. Bayesian measures of model complexity and fit. Journal of the Royal Statistical Society B 64:583โ639.