モデル選択手法の水産資源解析への応用 : 情報量規準とステップワイズ検定の取り扱い [in Japanese] Application of Methods for Model Selection to Fish Stock Analyses : Dealing with Information Criteria and Stepwise Tests [in Japanese]
Access this Article
Search this Article
Akaike's information criterion, which is widely used as a criterion of model selection in fish population dynamics, tend to overestimate the number of unknown parameters in several cases. In this paper, we discuss the selection performance to select the true model of various information criteria (AIC, BIC, c-AIC, TIC and HQ etc.) through computer simulations by analysis of variance, linear regression and Gaussian mixture model corresponding to CPUE standardization. As a result, we obtain the following results with regard to the goodness of various information criteria.<BR>1) In small samples or in the case that there are many unknown parameters compared to the sample size, the selection performance of c-AIC (finite correction of AIC) is superior to that of other information criteria.<BR>2) In large samples, the consistent information criteria, BIC and HQ, are better than AIC.<BR>3) In nested ANOVA-type model, the selection performance of TIC, which is exact evaluation of AIC, is almost same as that of AIC and c-AIC, and BIC is slightly good compared to stepwise F-test.<BR>4) In normal mixture model, stepwise chi-square test is not theoretically applicable and the selection performance of Bayes-type information criterion assuming the Dilichlet prior is superior to that of AIC and BIC.
- Japanese Journal of Biometrics
Japanese Journal of Biometrics 27(1), 55-67, 2006-06-30
The Biometric Society of Japan