The advent of high-speed computers is revolutionalizing many aspects of our life including ways of modeling our behavior. In this article, we overview some of the recent developments in Behaviormetrics triggered by the “computer revolution”, and suggest future developments, focussing on two topics of interest, nonlinearization and regularization in statistical models.
Behaviormetrics and Information Engineering/Science have been developing in parallel and different research fields, although they are both involved in “information processing”. However, the recent remarkable development of information processing and the communication infrastructure typical in computers and on the Internet is increasing the potential relationship between the two fields. In particular, the field of intelligent information processing necessitates more and more probabilistic and statistical methods as well as multivariate data analysis methods, an example of which is the expanding of its research fronts to deal with vast, various and uncertain information in the real world. The purpose of this article is to overview the recent research trends of intelligent information processing and to prospect the possibility and necessity of closer interaction and collaboration between behaviormetrics and information engineering in the future.
Rapid development in the survey methods available on the World Wide Web (WWW) is having a major impact on conventional survey data collection methods. The wide range of opinions has given rise to an ongoing debate regarding the future role of Internet surveys (in particular, Web surveys) based on the role that self-administration will play in research. We started by arranging a practical procedure for electronic data collection on the Web surveys experimentally designed from the viewpoint of “data science. Aiming to verify the applicability, possibilities, and limitations of Web survey methods, we conducted three experimental surveys during the period from 1997 to 2000. They were designed to enable comparison with each other and with traditional methods such as face-to-face interviews and online surveys using conventional sampling procedures. These surveys provided informative results about the characteristics of Web surveys. In the first survey, consisting of 12 continual surveys of a single panel of registrants, we examined the relationship between the response rates and the questionnaire's design, volume and content, as well as response rate differences among the 12 surveys and the discrepancies in repeated surveys. In the second experimental survey, we carried out Web surveys at about the same time on three different sites together with non-internet surveys using conventional sampling methods. Our experimental design enabled objective comparison of the surveys by using as much identical questionnaire design as possible. Our experimental surveys showed that Web survey results are similar to each other while distinctively differing from those of conventional surveys. In the third experimental survey, we simultaneously carried out a series of comparative surveys in order to examine the general characteristics of Web surveys found during the second experimental survey. Except that the number of sites used was two instead of three, the third experimental surveys were carried out in the same way as in the second trial. We confirmed the results that the same characteristics were evident again in the second survey. We also found that how the registrants of the surveys (named “resources”) were selected and whether the interval between solicitation and survey was short or long would be factors influencing the answers and response-rate. We also found that the respondents do not necessarily represent the resources. In addition, as an addendum in this paper, we report partly the results of a fourth experimental survey which has been carried out in 2001 to 2002 and compare it with the findings of the previous three trials. In particular, we also analyze the itemized causes of “nonresponse” on the datasets obtained from the tracking procedure of tracing electronically each respondent on the WWW. The fourth survey consists of Web surveys on three separate sites while the other surveys were based on conventional sampling methods (e.g., face-to-face interviews and mail surveys). While we use the same questionnaire design, content, and duration as those used in the past surveys, we also attempt to examine how the questionnaire design has influenced responses. Through these experimental surveys, an appropriate route to how to design a Web survey, evaluate its quality and avoid possible risks or perils in design is proposed from the concept of “data science.”
This paper is dedicated to the late Professor Kinji Mizuno in memory of his great contribution to Behaviormetrics and the Behaviormetric Society of Japan. In reviewing several recent issues on survey research in which Professor Mizuno was also very much involved, I consider possible future developments in this area. The issues cover the fields of public opinion survey, election forecast, Japanese national character survey and cross-national comparative survey by the Institute of Statistical Mathematics, as well as their methodologies. Finally, the idea of “the Advanced Institute of Human and Social Sciences”, an idea conceived by Professor Mizuno and his associates, is also briefly mentioned. I hope that this will remind our readers of the original concept of Behaviormetrics as it appeared in the 1970s and will lead us to the further development of our research.
Like a number of other sciences, the aims of Behaviormetrics are twofold. The first concern of Behaviormetrics is to clarify human behaviors by means of collecting data from human-related subject areas and analyzing them by mathematical methods. The second aim is to apply the new methods and knowledge obtained from research activities in Behaviormetrics to the development of human society. This article investigates how researchers of Behaviormetrics can contribute to human society in terms of learning from the essentials of Disaster Preparedness Education for School Children as studied by the late Professor Kinji Mizuno.
In the July 2001 election of the House of Councillors, THE YOMIURI SHIMBUN was able to conduct an accurate forecast of the landslide victory of the LDP in its nationwide survey on more than 60,000 eligible voters. For this poll, the YOMIURI adopted the telephone survey, selecting samples randomly from electoral registers. The suggested reasons for the gratifying results are the influence of “Koizumi phenomenon” and the improvements of the method of the survey. However, due to the large number of unlisted numbers, this survey actually failed to reach a considerable number of eligible voters. Furthermore, since listed numbers are already gradually on the decrease, by the next national election, the dilemma of reaching a significant number of voters will be even more problematic. Therefore, toward an accurate survey for the next election, improved methods of sampling are necessary, and in the case of the telephone survey, an adoption of RDD (Random Digit Dialing) is suggested.
The Mainichi Shimbun conducted a Telephone Survey before the 19th Upper House Election of Japan in July, 2001. In that Survey, a kind of media poll, two types of Telephone Survey Method were used, the Telephone Directory Method (the TD Method), and the Random Digit Sampling Method (the RDS Method). The TD Method was used to cover 32 electoral districts in rural areas and the RDS Method was employed for 15 districts in mega-city areas. When compared with the ballots of candidates of all districts in the Upper House Election, the polls of these two types of Telephone Survey Method and the regression analysis based on the polls had effectively corresponded to the results of the Japanese voters' choices. Nevertheless, considering that the precision of a survey depends on rigid examination, which involves the representativeness of the sample, valid response rates, and the characteristics of respondents, the TD Method has a vital problem impossible to eradicate, and the RDS Method has quite a few faults to correct, as well.
The Asahi Shimbun telephone survey made the transition from the list-assisted method to ASAHI-RDD (Random Digit Dialing) method in its survey prior to the Upper House Election in July, 2001. In fact, the Asahi Shimbun had estimated that more seats would be taken by the LDP than the actual results showed in the Upper House election in 1998 and in the Lower House election 2000, where they were without a win. The most probable cause of past failures consists of the non-coverage error since the list-assisted method had serious faults due to the large unlisted population. As RDD promised to give better results, Asahi decided to make use of Random Digit Dialing. While the Asahi method consists of Bank2 and random sampling, it does not use quota sampling. Its bank2-map includes the data, the number of listed-households and the overlap rate of area. A time schedule is created arranging the number of interviewers according to each day of the week. In addition to the problems caused by the current widespread use of mobile phones, there are many other problem areas which we must investigate and make efforts to solve.
Each month since 1994, JNN (Japan News Network) has conducted a nationwide public opinion survey on cabinet support and political party support, employing two different methods, the face-to-face interview survey and the telephone survey. This paper presents a comparison of the results derived from these two different suevey methods. The comparisons are made taking the following into consideration: the support rates of the present cabinet, the Liberal Democratic Party and the Democratic Party of Japan as well as the rate of the non-committed for each month. Our comparisons have found that there are, in fact, very large differences between the results in the interview surveys and those in the telephone surveys in the cases of both the support rates for the present cabinet and the rates of the non-committed for each time period. This indicates that different methods could cause different results even when we ask the same questions. On the other hand, when we see these results as time series data, we find a very high correlation (0.95) between the results of the two methods in the case of the rate of present cabinet support. However, in the other cases, the correlations are not as high as those of present cabinet support, and the size of the correlations seems to depend on the nature (e.g., the simplicity) of the questions.
In the statistical data analysis process, data investigation plays a most basic and important role. Its importance is emphasized and illustrated as the initial examination (Cox & Snell, 1981; Chatfield, 1985, 1991; Goto, 1986). In each step of data investigation, graphics is highlighted as one of the leading tools. However, in general, data investigation not only includes statistical graphics, but also formal analytical techniques. In the paper, we tried to conduct the data investigation based solely on the statistical graphical techniques. Then, as an elementary tool of graphical diagnosis, we proposed the data-adaptive probability plot. Moreover, we constructed a guardrail on the data-adaptive probability plot for inferential interpretation of the result of the plot. Furthermore, some useful aspects of the data-adaptive probability plot were assessed in several practical examples. As a consequence, we can say that the data-adaptive probability plot is a more flexible and interpretable tool than ordinary probability plots.