PROCEEDINGS OF THE FIRST US/JAPAN CONFERENCE ON THE FRONTIERS OF STATISTICAL MODELING

VOLUME 1, 2, AND 3

 

EDITED BY

H. BOZDOGAN

 

KLUWER ACADEMIC PUBLISHERS 1994

DORDRECHT, THE NETHERLANDS

ISBN 0-7923-2600-8 (Set of 3 Volumes)

Ordering Address:

 

® Designed by H. Bozdogan & Created by Hugh Bailey, 1991.

DEDICATED TO THE 65TH BIRTHDAY CELEBRATION OF PROFESSOR AKAIKE


"Professor Akaike is quietly assembling his own theory of statistical estimation based on entropy,
information and likelihood, centered around the Akaike's Information Criterion (AIC). This theory
is more likely to survive than most, being based on data and common sense rather than dogma."
 
Paraphrased from
L. A. Baxter
Journal of the Royal Statistical
Society, A, 1991, 154, Part 2,
pp. 356-366.
 

EDITOR'S GENERAL PREFACE

 
On May 24 to 29, 1992, The US/Japan Conference on the Frontiers of Statistical Modeling: An Informational Approach, was held at the Department of Statistics, University of Tennessee, Knoxville, to commemorate the sixty-fifth birthday of Professor Hirotugu Akaike and to honor him for his revolutionary contributions to modern statistical theory and practice. This conference was originally conceived while I was a Visiting Research Fellow and Visiting Associate Professor at the Institute of Statistical Mathematics, Tokyo, Japan during January-August 1988. A steering committee was formed with several colleagues from the Institute of Statistical Mathematics, and the initial plans of the conference were laid out at that time.
The proceedings of this conference are being published in three volumes under the names of the three scientific core areas:
 
I. Theory and Methodology of Time Series Analysis,
II. Multivariate Statistical Modeling, and
III. Engineering and Scientific Applications.
 
The conference theme was the implementation of statistical modeling through an informational approach to complex, real-world problems.
Over the past twenty years, modern statistical activity has been moving away from traditional formal methodologies based on test theory. The traditional methodologies based on narrowly specified hypotheses have become difficult in solving complex, large scale real-world problems, and are woefully inadequate to meet the challenges of our societal needs. Current efforts among the most sophisticated are to identify instantaneously the best fitting model among a class of competing models for a given complex data structure with computational ease, interpretability, and elegance using a new and novel informational approach. This richer approach links statistics as an experimental science with high speed computation and supercomputer technology, and bridges the conventional frequentist and the Bayesian schools of thought in statistics. With this new approach traditional table look up is eliminated. Total flexibility and versatility are provided to practitioners, researchers, and students in solving their statistical problems.
This new approach, was originally pioneered by Professor Hirotugu Akaike starting as early as 1971. Akaike (1981) writes:
 
"In 1970, I received an invitation to the Second International Symposium on Information Theory, which was to be held in Tsahkadsor, Armenia, USSR. At that time, I was interested in extending FPE to the determination of the number of factors in a factor analysis model, a statistical model originally developed in Psychology. However, it was not at all clear what the prediction error of this model was. The pressure of the impending deadline for the submission of the conference paper was increasing and this caused several weeks of sleepless nights.
On the morning of March 16, 1971, while I was taking a seat in a commuter train, I suddenly realized that the parameters of the factor analysis model were estimated by maximizing the likelihood and that the mean value of the logarithmus of the likelihood was connected with the Kullback-Leibler information number. This was the quantity that was to replace the mean squared error of prediction. A new measure of the badness of a statistical model with the parameters determined by the method of maximum likelihood was then defined by the formula
AIC=(-2)log(maximum likelihood) + 2 (number of parameters)."
 
The development and the introduction of Akaike's (1973) Information Criterion, "AIC", marked the beginning of the era of systematic approach to present-day statistics, namely model evaluation and selection. AIC is clearly one of the most interesting and important developments in the field of statistics in recent years. See, for example, Kotz and Johnson (Eds.) (1992, p. 599): Breakthroughs in Statistics Volume I: Foundations and Basic Theory, Springer-Verlag, New York.
This development represents an important advance over the conventional framework of statistics as developed by R. A. Fisher and many others following in his footsteps.
By combining ideas related to what is now called "predictive efficiency" with the notion of Kullback-Leibler information, Akaike arrived at AIC for evaluating alternative statistical models which cleverly combines a measure of goodness-of-fit of the model with a term relating to the number of parameters used to achieve that fit.
Akaike's AIC provides an answer to the question of how much improvement in fit should an additional parameter make before it is included in the model and on what scale should that improvement be measured?
Of course, important fundamental work like this answers some questions and raises many others. A by-product of Akaike's work is that he has directed the attention of other researchers to important problems. As a result, many important statistical modeling techniques have been developed in various fields of statistics, biomedicine, control theory, ecology, econometrics, engineering, psychometrics, sociology, and in many other fields. Voluminous research papers, books, research monographs, and Ph.D. theses have been written using AIC, including that of this editor. Further new model selection criteria have been developed based on Akaike's fundamental work using the fascinating informational, or entropic approach as the underlying sound analytical building block.
Professor Akaike has also been inspirational as far as encouraging us to have the courage to participate in consulting even when sometimes at first we may not feel up to the challenge. Presently his approach and philosophy to statistics have become a new and modern way of statistical activity and research with very successful industrial and scientific applications.
Therefore, the Profession of Statistics is greatly in debt to Akaike for his fundamental contribution in this new trend of informational approach to statistics, and for repeatedly calling our attention to the very important model evaluation and selection problem.
We proudly dedicate these three volumes of the Proceedings of the US/Japan Conference to Professor Akaike as our token gesture of deep appreciation of his outstanding contribution to the field of "Informational Modeling" on the occasion of celebration of his sixty-fifth birthday.
More than 100 participants from the US, Japan, Canada, Denmark, Germany, Israel, New Zealand, Switzerland, United Kingdom, and the Netherlands attended this important conference. About 40 very distinguished statisticians who made their contributions to these proceedings gave the invited lectures during a five day very intensive conference.
The conference was sponsored by the Department of Statistics, which is a part of the College of Business Administration, by the Office of the Dean of the College of Business Administration (CBA), and the Office of Research Administration, all at the University of Tennessee, Knoxville; and the Institute of Statistical Mathematics of Tokyo, Japan. The American Statistical Association (ASA) was co-sponsor along with Japan Statistical Society (JSS), Japan Society of Applied Statistics (JASAS), Classification Society of North America (CSNA), Psychometric Society, Engineering Physics and Mathematics Division of Oak Ridge National Laboratory, and well known professional organizations and research centers.
The result was a very pleasant and stimulating conference.
It is with great pleasure and gratitude that I acknowledge the help of the co-editors both from the US and Japan, the invited contributors, the referees from around the world, the chairs of the sessions, all the financial sponsors and the co-sponsors of the conference. I am deeply grateful to Dean C. Warren Neel of the CBA and Professor David Sylwester, the Head of the Statistics Department, for providing me the opportunities and the financial support to execute this prestigious conference. I am grateful to Professor Kunio Tanabe, my co-chair from Japan, and the Director of Prediction and Control at the Institute of Statistical Mathematics, Tokyo, Japan for helping me to successfully realize the conference from the beginning to the end, and also sponsoring all the Japanese researchers.
I wish to thank Professor Manny Parzen of the Department of Statistics at Texas A & M University who accepted to deliver the entertaining banquet speech on the evening of May 28, 1992 with his "change jokes", and "hammers and nails".
As always, there are key people behind the scenes in putting every successful conference. Without the dedication and help of these key individuals almost surely one can expect disasters in holding conferences. In my case, I had the personal privilege of working with Judy Snow, our Departmental Head Secretary, as my Conference Coordinator. The success of the conference was largely due to Judy's exemplary attitudes, congeniality, high professional standards and ability to relate to the diverse needs of the participants. Without my looking over her shoulder, Judy managed the conference single handedly on a daily basis and made all the arrangements. She made life easy so that I could spend most of my time in putting the scientific program together. We worked as a team. I specially thank Judy from the bottom of my heart for making the conference a success. I wish also to thank Alberta Randles, our Departmental Secretary, and eight Volunteer Graduate Students: Jim Minesky, Yuehui Fan, Paul Bowen, Gary Schneider, Martha Nunn, Rob Collignon, Frank Van Manen, and Chen Xi for their wonderful help in driving the vans, setting up the coffee-breaks, the bulletin boards, escorting the participants, and so forth, throughout the conference. I am grateful to all of them.
My colleagues in the Department of Statistics provided a congenial atmosphere and acted as my local organizers. I wish to thank every single one of them for their help and encouragement all the way through. I can still hear the voice of Dr. Richard Sanders shouting from the corridor : "Ham go home!" as I was working in the late hours. Dr. Mary Sue Younger chaired the local Organizing Committee, and helped to design several field trips and excursions with a variety of interests in mind to have a break and change the pace in the five-day intensive meeting. These included taking a dinner Cruise on the Star of Knoxville on the Tennessee River, hiking in the Great Smoky Mountains National Park, and dining in the picturesque "Old City" of Knoxville. Her help is also gratefully acknowledged. Pandora Riggs, Editor of Publications Center, and Designer Hugh Bailey at the University of Tennessee worked very hard with me in capturing my conceptualized ideas in preparing the design of the conference posters and the brochures. They are both acknowledged also.
I wish to thank Dr. Joseph (Joe) Johnson, the President, Professor William Snyder, Chancellor, Dr. Sheadrick Tillman, Assistant Vice Chancellor of Research, Dean C. Warren Neel, the Dean of CBA, and Dean Michael Stahl, Associate Dean of CBA, and Professor David Sylwester, Head of the Statistics Department of the University of Tennessee in participating in the opening welcome and greetings of the conference.
Dr. David Larner, Publisher of Science and Technology Division of Kluwer Academic Publishers in the Netherlands visited the University of Tennessee prior to the US/Japan Conference and during the five-day conference to discuss in detail the publishing plans and the format of the proceedings. These three volumes are the final product of his keen interest in the subject matter as a former Physics Professor himself. I am grateful to his continued support, interest, and persistence throughout the project. I am also grateful to Mr. John Martindale, Editor of the Science and Technology Division of Kluwer in the North American Continent, for his understanding, patience, and support in the final production of the three-set volume. Myself, and the co-editors thank both Dr. Larner and Mr. Martindale for bringing out these volumes.
Last but not least, I would like to thank my wife Nancy and two sons Bedros and Kirkor for their love and patience during the planning stages of the conference, and during the editing of these three volumes when I was not around to participate in family activities. I am grateful for their understanding and sacrifice.
It is hoped that the reader will find many interesting ideas and challenges in these resulting proceedings volumes, and that the contributions will stimulate further research in many complex, real-life applications of statistics in meeting our societal needs using the "Informational Modeling" approach.
 
 
Hamparsum Bozdogan
The University of Tennessee
 
Knoxville, July 1993
 
References
 
Akaike, H. (1973). Information Theory and an Extension of the Maximum Likelihood Principle, in Second International Symposium on Information Theory, B. N. Petrov and F. Csaki (Eds.), Budapest: Academiai Kiado, 267-281.
 
Akaike, H. (1981). A New Look at the Statistical Model Identification, in This Week's Citation Classic, Institute for Scientific Information, CC/Number 51, December 21.
 
Kotz, S. and Johnson, N. L. (Eds.) (1992). Breakthroughs in Statistics Volume I: Foundations and Basic Theory, Springer-Verlag, New York.
 

PREFACE TO VOLUME 1:
THEORY & METHODOLOGY TIME SERIES
ISBN 0-7923-2597-4 (Volume 1)
 
This volume contains the papers of The US/Japan Conference on the Frontiers of Statistical Modeling: An Informational Approach which deal with the theory and methodology of time series analysis.
The first paper is the banquet talk given by Professor Parzen on the evening of May 28, 1992 and discusses Professor Akaike's career in statistical science, and the phases of his research career under the title: "Hirotugu Akaike, Statistical Scientist".
The second paper is one of the keynote lectures given by Professor Akaike and discusses his own experiences on the development of time series models. He gives a very interesting mix of real-life examples from his own career as a statistician. The rest of the papers included in this volume discuss problems of state space modeling of time series, autoregressive model fitting, system analysis and seasonal adjustment through model fitting, and dynamic analysis of Japan's economic structure, to mention a few. Interspersed in this volume are the papers which deal with applications of model selection in small sample times series problems, temporal causality, detecting changes in trend for short time series, harmonic non-linear regression, and applications of the TIMSAC (TIMe Series Analysis and Control) software package.
It is with great pleasure and gratitude that the editor and co-editors acknowledge the authors and the referees who contributed to this volume, and the chairs of this scientific core of the conference.
 
Hamparsum Bozdogan
The University of Tennessee
Knoxville, July 1993

CONTENTS OF VOLUME 1
 List of Contributors to Volume 1 ix
 Editor's General Preface xi
 Preface to Volume 1 xv
Summary of Contributed Papers to Volume 1
1. Hirotugu Akaike,Statistical Scientist
E.Parzen
 25
2. Experiences on the Development of Time Series Models (Keynote lecture) H.Akaike
33
3. State Space Modeling of Time Series
G.Kitagawa
 43
4. Autoregressive Model Fitting and Windows
M.B.Priestley
 63
5. System Analysis and Seasonal Adjustment Through Model Fitting
M.Ishiguro
 79
6. Akaike's Approach Can Yield Consistent Order Determination
H.Tong
 93
7. Recursive Order Selection for an ARMA Process
R.J.Bhansali
 105
8. Autoregressive Model Selection in Small Samples Using a Bias-Corrected Version of AIC
C.M.Hurvich and C.L.Tsai
 137
9. Temporal Causality Measures Based on AIC
W.Polasek
 159
10. An Automated Robust Method for Estimating Trend and Detecting Changes in Trend for Short Time Series
T.Atilgan
 169
11. Model Selection in Harmonic Non-Linear Regression
D.Haughton,J.Haughton,and A.Izenman
 187
12. Dynamic Analysis of Japan's Economic Structure
S.Naniwa
 209
13. New Estimates of the Autocorrelation Coefficients of Stationary Sequences
S.Batalama and D.Kazakos
 233
14. Applications of TIMSAC
Y.Tamura
 251