It is a great honour and pleasure for me to have been elected to the President of the JSS and to deliver the presidential address here today. Especially this year happens to be the 50-th anniversary of the JSS. The JSS has been playing a central role in promoting the statistical activities in Japan and I realize that the responsibilities of the President of the JSS have become heavier than before now. I can only say that I will try my best in such a way that your trust put upon me is not so grotesquely misplaced.
Now since I myself have been puzzled by The Fisher Randomization from the very beginning when I started to learn the modern statistics around 1944 and I have been working on the subject since 1959, I thought that it is an appropriate topic on which I shall talk about as my presidential address.
My talk consists of two parts: Part I will be devoted to a historical perspective in which I shall summarize what happened about the randomization as I see it and clarify my own standpoint about the randomization. Part II will be devoted to “The Fisher-Student Controversy on The Randomization”. There has never been given a comprehensive settlement of the famous controversy, so far as I am aware of. I hope that I can give a comprehensive settlement of the controversy based on the Barbacki-Fisher paper [10] from my own standpoint.

View full abstract

The purpose of this paper is to consider goodness-of-fit statistics for the case that the underlying distribution is previously known to be symmetric about zero or unknown location parameter and derive the asymptotic distributions of these statistics. In addition to this, a statistic for testing normality is introduced and Monte-Carlo study of power comparisons is performed.

View full abstract

We consider the goodness-of-fit tests based on the sample spacings. In this paper, the test statistic T_n=n^-1_??_(G_??_-m/(n+1))² is introduced in terms of high-order spacings G_??_, and it is shown that the test based on T_n is asymptotically more powerful than the previous tests based on the sample spacings and that T_n is asymptotically equivalent to the Watson statistic.

View full abstract

The problem of selecting variables in normal linear regression model is considered as a multiple decision problem. A loss function is introduced which is zero if the true model is selected and one, otherwise. Then some variable selection rules proposed by several authors are shown to be admissible.

View full abstract

As early as the first paper of Bechhofer on ranking and selection in 1954 it was recognized that, with a ranking and selection goal as well as with other goals such as estimation or hypothesis testing, it might be desirable to carry out one's experiment in some design other than the completely randomized design. Nevertheless, over the years almost all of the papers in the area have developed their methods and theory explicitly only for the completely randomized design. In this paper we review what is known about ranking and selection in design settings other than the completely randomized design, and then proceed to new results on the complete factorial experiment setting.

View full abstract

The maximum likelihood estimation of two-phase segmented lines for binary data is considered. The asymptotic distribution of the maximum likelihood estimators for the parameter is obtained under regularity conditions. It is also demonstrated that the sequence of the maximum likelihood estimators converges to the true value of the parameter with probability 1 under the same regularity conditions. As an example two-phase lines are fitted to dose-response relations revealed on radiation exposure data.

View full abstract