Behaviormetrika
Online ISSN : 1349-6964
Print ISSN : 0385-7417
ISSN-L : 0385-7417
23 巻 , 2 号

• Hisao Miyano
1996 年 23 巻 2 号 p. 129-139
発行日: 1996年
公開日: 2006/05/19
ジャーナル フリー
In this paper, we propose a new clustering method based on the concept of maximum likelihood (ML) estimation. In general, the problem of local minima arises when we try to use the ML method in clustering problems. Our method circumvents this problem by employing the so called simulated annealing technique. In section 2, we formulate our clustering problem using the ML concept, and derive the ML estimation method. In section 3, validity of the derived method is confirmed by analyzing two artificial data and the famous Iris data. In the final section, our method is also extended from the viewpoint of sequential estimation.
• Nobuoki Eshima
1996 年 23 巻 2 号 p. 141-152
発行日: 1996年
公開日: 2006/05/19
ジャーナル フリー
This paper compares an association model for trivariate contingency tables with the trivariate normal distribution. First, similarity between the two models is discussed. Second, it is proved that the association model approximates the discretized trivariate normal distribution. An aritificial illustration shows the degrees of the approximations in small tables. Third, estimation of the correlation coefficients of the underlying trivariate normal distribution is discussed. The present approach is illustrated by use of numerical examples.
• Yoshio Takane
1996 年 23 巻 2 号 p. 153-167
発行日: 1996年
公開日: 2006/05/19
ジャーナル フリー
An item response model, similar to that in test theory, was proposed for multiplechoice questionaire data. In this model both subjects and item categories are represented as points in a multidimensional euclidean space. The probability of a particular subject choosing a particular item category is stated as a decreasing function of the distance between the subject point and the item category point. The subject point is assumed to follow a certain distribution, and is then integrated out to derive marginal probabilities of response patterns. A marginal maximum likelihood (MML) method was developed to estimate coordinates of the item category points as well as distributional properties of the subject point. Bock and Aitkin's EM algorithm was adapted to the MML estimation of the proposed model. Examples were given to illustrate the method, which we call MAXMC.