Bulletin of the Computational Statistics of Japan

We have developed a multiple regression analysis supporting system named RASS2. The aim of the system is to provide consulting mechanisms with high flexibility and extendibility to users. RASS2 is constructed with original objects as its basic components. Objects in RASS2 are extended objects, which include rules for representing knowledge of regression analysis techniques. Rules are described within the class definition. Each object suggests the best procedure according to the state of the analysis and negotiates with other object to resolve conflicts among their suggestions. Users can follow the suggestion or can execute other procedures. Compared with the earlier version of RASS, the objects make the system structure simple and modifications of statistical knowledge embedded in the system easy.

View full abstract

Show abstractHide abstract

In this paper, we discuss a parallel execution about data analysis. Especially, we focus on the execution in parallel virtual computer consisted of many micro computers connected each other via network, because we can get this environment more easily than we get a super-computer. We have some simulations with the Bootstrap on PVM (Parallel Virtual Machine) environment, which is one of the popular software libraries to have some computers the virtual one, in order to investigate characteristics on execution time mainly. The parameters in the simulations are the number of Boostrap samples, the number of processors, calculation load in slave (expressed in a waiting time). The results show some important features and the relation between parameters and a round trip time, and we get a simple model with regression. This model is effective to consider execution time from an analysis on a virtual machine or how many personal computers we should prepare to analyze data in time, etc.

View full abstract

Download PDF (1050K)

In this paper the author has investigated the numerical experiments and illustrated examples of the application of feed-forward neural networks to regression analysis with qualitative variables. The author numerically illustrates the effectiveness of the feed-forward neural network models with appropriate variables related to the qualitative variables. In this model, the optimal numbers of hidden layer units decided by the AIC under the condition that such factors had not been directly observed. The comparative studies have shown the proposed approach effectively predicts the momentary volume flowing water only with an appropriate auxiliary variable as input variable to the neural networks.

View full abstract

In psychometrics, some methods of multivariate data analysis have originated and developed according to the needs in psychology. Historical and recent developments of such psychometric methods are reviewed. Following the summary of historically important methods (Thurstonian scaling and classical test theory), we review descriptive multidimensional analysis (optimal quantification, extended principal component analysis and multidimensional scaling) and latent variable analysis (factor analysis, structural equation modeling and item response theory). Additionally, we mention two related fields, mathematical psychology and statistical education in psychology.

View full abstract

Register with J-STAGE for free!