2012 年 76 巻 3 号 p. 115-122
To discuss the problem of estimating age composition from length-frequency data in a finite normal mixture distribution, I introduce a hybrid method for estimating the number of components in the full model without limitation and a structured model with relationship among unknown parameters. The computational procedure was expressed in four steps and statistical interpretation/significance of the hybrid method suggested by Eguchi and Yoshioka in 2001. The hybrid method was applied to the well-known length-frequency data for yellow sea bream, Dentex hypselosomus. The computer simulation experiment estimated the number of components in the finite normal mixture model whose probability density function resembles a log-normal distribution. Results showed that the selection performance of the hybrid method for choosing the true model was superior to traditional methods such as Akaike’s information criterion (AIC) and/or Baysian information criterion (BIC).
