The ASA Statistical Computing Section
A History
John M. Chambers and Barbara F. Ryan
The involvement of statisticians with computing goes back a very long
time, long before any device existed that we would call a computer
today. The U.S. Bureau of the Census and its predecessors, for
example, have wrestled since the inception of the decennial census
with the challenge of recording, transmitting, storing, and processing
the growing volume of data (e.g., see a fascinating history by
K. S. Reid-Green in the February 1989 issue of Scientific
American). Since American Statistical Association (ASA) members have
always been instrumental in census activities, we can confidently
assert that ASA has been concerned with computing from the
association's earliest years. Similar arguments apply also to the
statistical and computing needs of other organizations, such as the
U.S. Bureau of Standards.
With the advent of programmed computing on a substantial scale in the
1950s and 1960s, the involvement of statisticians in research,
development, and use of computing facilities became increasingly
central to much activity in the profession. Growing recognition that
the ASA should acknowledge the role of computing culminated in the
formation of the Statistical Computing Section in 1972 and in the
section's contributions to the association and the profession since
then.
By the middle of the 1960s, much statistical software existed, of
varied styles and aimed at varying applications. Early software may
seem crude from our present perspective, just as the computers it ran
on seem amazingly primitive now. But these early efforts had already
changed statisticians' views of their activities profoundly and
permanently. Those who were involved in statistical computing had a
conviction that associations like ASA had to respond by recognizing
the new field. Several events and activities during the 1960s provided
special impetus to that conviction and the formation of the section.
In December 1966 a conference on statistical computing was held in the
United Kingdom, under the auspices of the Royal Statistical
Society. Papers at the conference presented both existing software
systems and general ideas for future directions. The papers appeared
later in the journal Applied Statistics. Discussions during
and after the conference led to the formation of a Working Party on
Statistical Computing by the Royal Statistical Society. Though the
focus of the conference was naturally British, one of the papers and
several of the participants in the conference came from the United
States. Some of the participants, including John Chambers and Paul
Meier, returned later to the United States, anxious to see ASA become
involved in this area as well.
Beginning in 1968, a sequence of meetings generally called the
Interface Symposia created a remarkable self-sustaining forum for
exploring the interactions between statistics and
computing. Originating in California from the efforts of Arnold
Goodman and Nancy Mann, the symposia have flourished ever since and
continue to draw together the two disciplines. In 1969 another
conference was held on statistical computing, this time at the
University of Wisconsin-Madison. Papers from this conference were
published as a book. Statistical Computation, edited by Roy
Milton and John Nelder.
The association responded to all of these stimuli at a typically
deliberate pace. The first concrete step was the appointment of a
Committee on Computers in Statistics, originally appointed to serve
for 1970. The committee had 10 members: Meier (chair), Frank Anscombe,
Joe Cameron, Chambers, Joe Daly, Art Dempster, Wil Dixon, Michael
Godfrey, Merv Muller, and Martin Schatzoff. As was typical of all
early ASA computing activities, the members ranged from those directly
involved in computing research to sympathetic observers who rarely, if
ever, touched a keypunch themselves. The name given to the committee
suggests, in retrospect, a continuing emphasis on computing machinery
as something with an impact on statisticians, rather than on computing
as an activity by statisticians. Change, however, was clearly on the
way. The committee sponsored three sessions at the December 1970
annual meeting, marking the first time that computing had an official
part at the meeting. The sessions were "Algorithms for Statistical
Computation," "Computers in Teaching Statistics," and "Data Handling."
The next major step also came out of the December 1970 meeting: a
petition with 75 signatures calling for an ASA section on statistical
computing. Dixon sent the petition to Churchill Eisenhart, ASA
president, on January 5, 1971. This event finally set in motion the
creation of the section. Eisenhart replied sympathetically, outlining
the steps to be taken to get approval and asking Dixon to prepare a
charter statement. That draft charter cited six purposes for the
section. The first and the sixth defined desirable interactions, in
both directions, between statistics and computing: "to encourage the
interest and involvement of the membership of ASA in statistical
computing" and "to encourage the applications of statistics in the
design, maintenance, and evaluation of computing hardware and software
systems." The first has been the main theme of the section ever
since. The other, the application of statistics to computing, has
flickered in and out over the years in the form of compumetrics,
program evaluation, software quality, and other topics. The other four
purposes involved activities such as interacting with other
organizations and providing a "resource" in statistical computing.
The approval process went forward smoothly, and a section was approved
at the 1971 annual meeting. The first officers, appointed for 1972,
were Dixon as chair, Dempster as program chair, and Alan Forsythe as
secretary/treasurer. In its first election the section chose a partial
slate of officers (Chambers as chair and Wes Nicholson as program
chair) for 1973 and a full set of officers for 1974. From then on, the
section was in business.
As the new field evolved, the section evolved with it. In the early
1970s statistical computing was still working toward a clear identity
and recognized respectability, especially among the more
mathematically inclined. Theorems, or even crisp results, are rare,
and some of the most important developments, such as the interactive
exploration of data, cannot be adequately captured in written form.
The computing-section sessions at the ASA annual meeting and the
annual Symposium on the Interface were the primary homes for
statisticians interested in computing. Even though the formal ties
between these two groups were loose, the informal ties were strong and
close; we regarded the two meetings as our spring and summer meetings
on statistical computing. Journals were not readily available to us,
too slow for the fast-moving technology, and often inappropriate for
our work. So we exchanged our ideas and software informally, at these
meetings.
Certain basic areas in computing emerged early—simulation,
algorithms, computerized data analysis, and software. These remain the
major areas in statistical computing today.
Surprisingly, perhaps, statisticians do not do much basic research in
simulation as a technique, but they do use it extensively. Most
theoretical work is supported to some extent by simulation, and most
ASA meetings contain talks that use simulation. A few meetings,
including the first for this section, contain sessions that develop
simulation methodology. In 1985 the section became a cosponsor of the
Winter Simulation Conference.
Algorithms, on the other hand, have the largest representation of the
four areas mentioned previously. They form the supporting base for
computing and hence for most data analysis. In the 1970s, research in
the area emphasized algorithms for linear models, primarily analysis
of variance and regression. Most ASA meetings had at least one
invitedpaper session on this topic, and they always had many
contributed papers. We were all trying to learn how to do calculations
efficiently and correctly, how numerical analysts could help us, and
how calculations were actually being done in commercial software.
In the 1980s, work in algorithms was much more diverse and included
traditional research in linear methods as well as the new
computer-intensive methods for data analysis. The annual meetings had
talks on mixed models in analysis of variance (ANOVA), optimality
criteria for designed experiments, projection pursuit, nonparametric
regression, clustering, methods for analyzing incomplete data, image
analysis, Bayesian methods, and smoothing.
In 1981 the Committee on Statistical Algorithms was formed, with Sally
Howe as chair, to develop a classification scheme for statistical
algorithms and an index of available algorithms. During the next few
years the committee compiled a guide with more than 2,500 entries.
We also explored ways to use computers effectively to analyze
data. The area discussed most, with a total of seven sessions, was
data management, including how to manage large, complex data sets, how
to assure the quality of research data, and how to store data
efficiently. Some of the most popular sessions were on the use of
computers for graphics, from simple line-printer plots to interactive,
dynamic graphics systems. In addition, there were talks on using
computers in specific areas of analysis, including biased estimation,
missing-data analysis, survival analysis, regression, ANOVA and the
analysis of frequency data.
The last major area in computing, software, contains some of the most
interesting and controversial talks, those that evaluated statistical
packages. The 1970s was the era of the central mainframe computer, and
most statistical packages were large batch programs (remember punched
cards). User interfaces were unfriendly, the methods were often
obscure, and the answers were sometimes even wrong.
In 1973 Ivor Francis and Richard Heiberger asked the section to form a
committee to evaluate statistical packages. With Paul Velleman, they
put together a report covering criteria, considerations, and test
plans for evaluation. The report was presented at the 1974 ASA annual
meeting and subsequently appeared in The American
Statistician. During the next eight years the Committee on the
Evaluation of Statistical Program packages and its successor, the
Committee on the Evaluation of Statistical Software, organized more
than a dozen sessions at the interface and ASA meetings, evaluating
various aspects of large batch packages, such as BMDP, Omnitab,
P-Stat, SAS, and SPSS. There were talks on regression, ANOVA, cluster
analysis, time series, nonlinear regression, graphics, and user
documentation.
In 1983 Ken Berk, then chair of the committee, wrote a letter to the
editor of The American Statistician, suggesting that software
reviews appear in print so that more people could use them. The next
year, he was appointed editor of a new section of The American
Statistician-Statistical Computing Software Reviews. In February
1985 the first reviews were published.
The lack of appropriate publication outlets for work in statistical
computing has been a problem and topic of discussion since the
beginning of the section. As early as 1973 we explored the possibility
of creating a journal appropriate for our work. As a first step, in
1975 the section started publishing the proceedings of papers
presented at the ASA annual meeting. In December 1988 William Eddy, as
editor, with six associate editors, put together a special
Applications section in the Journal of the American Statistical
Association. Active promotion of a new journal continues.
The section is governed by an executive committee consisting of a
chair, chair-elect, program chair, program chairelect,
secretary/treasurer, board representative, publications liaison
officer, two American Federation of Information Processing Societies
representatives, and a representative to the committee on
sections. The committee meets at the ASA annual meeting and, since
1981, also meets at the interface meetings. In 1983 Barbara Ryan
started what has become a tradition of having a mixer, sponsored by
commercial software developers, after the section business meeting.
|
While writing this article, John M. Chambers
was Statistician, Statistics Research Departments, AT&T Bell
Laboratories, Murray Hill, NJ 07974. He is still with Bell
Laboratories, but Bell Labs is now part of Lucent Technologies.
Barbara F. Ryan was a Visiting Scholar,
Department of Statistics, Stanford University, Stanford, CA 94305. She
is now President and Chief Executive Officer of Minitab, Inc.
This paper originally appered in The American Statistician,
May 1990 (Volume 4, No. 2, pp 87-89).
|