StatisticalComputing TeachingStatisticsand ...stat-computing.org/newsletter/issues/scgn-05-3.pdffor obtaining assistance for learning and practicing sta-tistical methodology. ... [email protected]

A joint newsletter of theStatistical Computing& Statistical GraphicsSections of the AmericanStatistical Association.

December 94 Vol.5 No.3

A WORD FROM OUR CHAIRS

Statistical ComputingOur thanks to Trevor Hastie, theoutgoing 1994 Chair of the Statisti-cal Computing Section. He recentlymoved to Stanford University, Depart-ment of Statistics, from Bell Labora-tories.

It was nice to see our sessions at the ASA meetings inToronto so well attended. I counted over 250 peoplein the neural network session featuring Geoff Hinton.Our business meeting and mixer, joint with graphics,was also a great success. I think the fact that we haveimproved the spread over the years has helped!

Let me summarize the highlights of the executive meet-ing of our section, which took place during the jointstatistical meetings. One of our main missions is topromote statistical computing to students and new ASAmembers. CONTINUED ON PAGE 4

Statistical Graphics

Roy Welsch is the outgoing 1994Chair of the Statistical Graphics Sec-tion. The editors wish to thank himfor his leadership and help in provid-ing this column during the past year.

It seems like just yesterday that I was writing my firstcolumn as Chair of the Statistical Graphics Section.Given my record at meeting Newsletter deadlines itprobably was just yesterday! The time is here (andthe deadline has passed) to write my final column be-fore giving the pen to David Scott, next year’s Chair.Next year I, too, will be able to enjoy my Thanksgivingvacation. CONTINUED ON PAGE 2

FEATURE ARTICLE

Teaching Statistics andComputing via Multimediathrough the World WideWebby A.J. Rossini and J.L. Rosenberger

SummaryHypertext is growing in use as a method for informationretrieval and dissemination. The introduction of Mosaichas led to a rapid expansion of information services onthe Internet. We propose the abstraction of statisti-cal data, information, and procedures for disseminationand teaching through this medium. The power of hy-pertext unleashes a new approach for constructing anddistributing learning materials for statistical courses. Aparadigm of statistical viewers is discussed, and an im-plementation using Xlisp-Stat and Hyper-Text MarkupLanguage (HTML) is described for UNIX and the X11Window System. Technical details are presented, anda demonstration is available on-line on the Internet toexperience first hand the potential of this medium.

IntroductionIn attempting to understand statistics, a typical studentspends the majority of time outside the classroom learn-ing and studying the material. The popularization ofthe Internet provides students with an additional meansfor obtaining assistance for learning and practicing sta-tistical methodology. This article describes a varietyof instructional mechanisms using the World Wide Web(WWW ), a network or virtual web of documents and in-formation sources accessible from any computer on theInternet world-wide. Mechanisms are described andexamples given for a variety of possible instructionaluses. CONTINUED ON PAGE 10

EDITORIAL

This December 94 issue (Vol. 5, No. 3) of the Newslet-ter is the last issue for many of the current section offi-cers. As the new officers take their positions, you shouldbe thinking about how you might contribute to our sec-tions. Also we challenge you to suggest colleagues whomay be able to contribute for their own professional de-velopment, and to enhance the sections’ activities. JohnRice, program chair, is requesting volunteers for chair-ing statistical computing sessions at the 1995 OrlandoMeeting of the ASA. Please contact him if you wish tonominate a junior colleague, or if you wish volunteer tochair a session yourself.

Finally, we are still looking for suggestions of personswho would like to edit this Newsletter for the next 2–3year term. It is not a lonely task—the sections havebeen cooperating—with a co-editor from each section.

This issue continues the recent multimedia blitz aboutthe multitude of uses of the Internet. Our feature articleby Rossini and Rosenberger is but one example of sucharticles, which describes an application of the Internetwhich should change the face of distance education for-ever. Syllabus magazine also featured "Education andthe Internet" in their Nov/Dec 1994 issue. Although weare sensitive to the need for balance in this Newsletter,these two articles on XlispStat converged at the sametime, and seemed to fit together nicely. Since XlispStatis in the public domain, we hope the attention promotesthe profession, and not just the developers.

The Statistical Computing Section of the ASA has ap-proved a student paper competition for the annual jointstatistical meetings, to be held in Orlando next August.Advisors should encourage their graduate students tosubmit an abstract and paper on a topic of interest to thestatistical computing section. The winners will receiveregistration and travel funds, and the rest will still havethe opportunity to enter their paper as a contributed pa-per or poster session—winners all. If you are unable tomake the deadline for this year, then this reminder is forlong-range planning for next year, since the intention isto make this an annual event. See details in our chair’scolumn and on page 5.

Our regular columnists have continued their run of inter-esting articles. Mike Conlon discusses the issues arisingin establishing a computer network server in his Depart-mental Computing column; Dan Carr investigates thegraphical issues involved in producing boxplots, witha redesign of several features and the introduction ofcolor to assist the eye in making comparisons. Is thisevolutional shift an improvement? I’m sure he would

appreciate feedback, and we will gladly publish any let-ter to the editor on the topic. In the Net Snooping col-umn, Mike Meyer describes a new product, Netscape,which has challenged Mosaic as the interface of choiceto the Internet. The landscape is moving very quickly,and it should be fun to watch the rapidly emerging prod-ucts available as the marketplace pushes it’s way intocyberspace.

We again wish to encourage submissions of articles tothis newsletter. Letters to the editors, short informativepieces, information, and longer articles are all welcome.Send submissions and questions about appropriatenessvia e-mail to the editors. We encourage submissions inLATEX or TEX, but will gladly accept plain ASCII textfiles as well.

James L. RosenbergerEditor, Statistical Computing [email protected]

Mike M. MeyerEditor, Statistical Graphics [email protected]

FROM OUR CHAIRS (Cont.): : :

Statistical GraphicsCONTINUED FROM PAGE 1

I do have something to be thankful for and that isall the help that I have received as from many vol-unteers. Linda Clark has superbly managed our Sectionfinances (without derivatives!), created logical minutesfrom meetings where the discussion was fast and fu-rious, and co-produced the joint mixer at the AnnualMeetings in Toronto.

Bill DuMouchel put together the excellent program thisyear while moving from Boston to New York to takea new job. Rick Becker has been a fountainhead ofknowledge as past chair and was always reminding meof the things I needed to do to carry on the initiativesthat he started. (Got that, David.)

Jane Gentleman and Sally Morton have carried ourviews to the Council of Sections and endured the or-ganizational machinations of the ASA. Sally also con-tinued her excellent management of our video libraryand volunteered to take over as Section Program Chair-Elect from Sallie Keller-McNulty when Sallie needed to

2 Statistical Computing & Statistical Graphics Newsletter December 94

resign because of her new responsibilities at NSF. Sallieand Sally have already assembled an exciting programfor Orlando. Deborah Swayne will assume the videolibrary position so that Sally Morton will have time forsomething other than Section business. Although nota Section officer, Lorraine Denby is always ready tohelp and has informally served as Section archivist andhistorian while encouraging us to strike out in new direc-tions as well. Lorraine also becomes the new member ofthe Section Committee to Nominate Fellows replacingLuke Tierney. Congratulations to Andreas Buja whowas nominated by the Graphics Section and became aFellow this year.

Jerry Moreno deserves to be thanked again for manag-ing the very successful Graphics Poster Competitions.Please help to continue the tradition by saying yes whenJohn Schollenberger, Jerry’s successor, asks for assis-tance. Robin Lock will be handling the Data Expositionfor the Orlando Meeting. He would welcome sugges-tions for data sets to be used in the Exposition.

Thanks also to Abbe Hertzig who is stepping downas head of continuing education for the Section. Herreplacement, Karen Kafadar, has been busy and has atleast one session ready for Orlando entitled "Using tcl/tkin building interfaces to statistical software and robustvisualization" which will be given by Joseph McKeanand John Kapenga from Western Michigan University.

Mike Meyer deserves everyone’s praise (even those ofus with deadlines) for co-editing (with Jim Rosenbergerfrom the Statistical Computing Section) the Newsletteryou are reading. Mike is finishing his term as editor andhe and the Section are looking for suggestions for a neweditor.

Our strong budget position has allowed us to take sev-eral new initiatives. We will be funding more outsidespeakers at our meetings and providing them with firstclass equipment to give their presentations. To pro-vide an incentive for authors to send their best workon graphics to the Journal of Statistical Computing andGraphics (JCGS), I have appointed a committee consist-ing of Rick Becker (Chair), Howard Wainer, and JerryMoreno charged with setting up a process that will leadto the awarding of a prize or prizes for the best paperspublished in JCGS. This committee may also be askedto consider an evaluation process and prizes for the bestgraphics presentation at the annual meetings.

We also will be spending more to promote Section activ-ities, to subsidize continuing education when necessary,to provide incentives for Section members (reduced feesand subscription charges), and to encourage student par-ticipation at all levels. This year the Section supported

a Data Analysis Contest for undergraduate mathemat-ics and statistics majors run by David Robinson at St.Cloud University. We also want to help the ASA andall Sections get involved the Internet, Mosaic, and theWorld Wide Web since these tools can now be used toprovide graphical communication. This may requiresome funds as well.

Rick Becker as Chair of the Nominations Committeehas proposed the following candidates for the 1995 Sta-tistical Graphics Election:

Chair-Elect: Sally Morton and Linda Clark

Program Chair-Elect: Robert Wesley and Di Cook

Council of Sections Representative (96-98): LorraineDenby and Jerry Moreno

Final notes:

Don’t forget the Interface ’95 Symposium in Pittsburgh,PA June 21-24, 1995.

Please give some time and energy to your Section. Thatis what makes things happen.

My apologies to David Scott for not returning his blacktransparency pen that I borrowed to make some slidesfor the Section business meeting/mixer. David men-tioned this at his talk when he had only slides madewith a red pen. At least I used to prepare my talks onthe plane coming to the meeting, not at the meeting!

It is now a pleasure (and relief) to introduce David Scottto this column. Good luck, David!

Roy WelschChair, Statistical Graphics [email protected]

Call for IdeasThe Statistical Graphics Section officers are interestedin hearing from members about effective and appropri-ate ideas for budgeting and expending section funds.General or specific ideas should be sent to David Scott,Chair-elect, [email protected].

David ScottChair-Elect, Statistical Graphics SectionRice [email protected]

December 94 Statistical Computing & Statistical Graphics Newsletter 3

FROM OUR CHAIRS (Cont.): : :

Statistical ComputingCONTINUED FROM PAGE 1

Our Continuing Education representative, Tom Devlin,proposed to enhance the scholarships given by our sec-tion for students to attend short courses co-sponsoredby our section. These courses are given at both the In-terface meetings and the ASA. The executive approvedhis proposal, and the scholarships will cover both thecourse and conference registration fees for up to 20 stu-dents per year. Details on how to apply will be providedin the next Newsletter. We also voted to augment thesupport of presenters of short course sponsored by oursection at the ASA meetings. Currently they receivean honorarium of $500 and expenses of $1000 towardstravel and lodging, from the ASA, to be shared by allthe presenters. Our section will augment this by $250per presenter, for up to three presenters.

The deadline for the student paper competition is ap-proaching (January 10, 1995). This is a wonderful op-portunity for students to attend the ASA meetings fullyfunded, give a paper in a special contributed paper ses-sion, and have their paper summarized in this newslet-ter! Details appeared in the last newsletter, August 94,page 3, and is reprinted here on page 5.

Leo Breiman of U.C Berkeley is giving a tutorial onstatistical modeling at the 1994 Neural Information Pro-cessing Systems (NIPS) conference in Denver this De-cember. This is the flagship conference for Neural Net-works, and our section is co-sponsoring his tutorial.Other section members are participating at this meetingas well: Rob Tibshirani, Jerome Friedman and I are allgiving papers.

Statistical Computing IssuesNow to an issue that has been on my mind lately: STA-TISTICAL COMPUTING! Statistics has certainlybeen changing over the years, and by far the biggestimpact has been the effect of computing speed, datatransmission and storage capacity. Think of some ofthe hot topics in the field right now: the bootstrap andresampling schemes for inference; Gibbs sampling inBayesian posterior computations; maximum-likelihoodestimation of complicated and highly connected geneticmodels; likewise for graphical Bayesian belief mod-els; wavelet modeling of high dimensional signals andimages; statistical models where observations are im-

ages; neural network and other non-parametric mod-els for classification and regression with huge trainingsets; high dimensional inverse problems (PET); pat-tern recognition in high-resolution satellite images; realtime (interactive) statistical graphics. In some cases(e.g. bootstrap, Gibbs) these techniques are supplantingsome of our traditional statistical and probability mod-els, created for mathematical and computational conve-nience but relying on unjustified assumptions. In othercases (e.g. images, interactive graphics) we can dabblein areas that were simply not available to us in the past.These are all computer intensive areas, and require sig-nificant resources and computing skills. One reflectionof these changes has been the recent burst of new statis-tical computing journals—it has become legitimate topublish computing type articles (and get credit from ourpeers for doing so!)

What about our computing needs, in statistics depart-ments at universities and other research centers? Weare beyond the age of the “package” that can do allthings statistical. We need well integrated and plannedhardware and software environments, and facilities andabilities for programming at a variety of levels. Someof the areas require really specialized equipment andsoftware. Most important, these are not one time needsthat a few committee meetings and a lump of moneycan solve.

Here comes a change that statistics departments havethe most difficulty dealing with. With all these com-puting demands, a Computer Resource Administrator isneeded. This is a specialist who can manage the plan-ning of the environment; who manages software andhardware acquisition and maintenance; who can takenotice of and try out innovative software and freeware;who can teach introductory and not so introductory com-puting classes to the faculty and graduate students; whocan manage data bases, create WWW resources andmanage ftp sites. The list goes on.

Why is it so hard to justify such a person? Mainly be-cause it sounds like a faculty position will have to besacrificed, and they are like hen’s teeth around univer-sities. In many cases the people making these kindsof decisions have not been swept along by the tides ofchange, and don’t feel the need as strongly.

I recently conducted a mini-survey of departments whohad managed in some form or another to create such aposition. Create is the word: in a large number of casesthe University and department together imaginativelyconjured up a way to fund the positions from grants,department funds and campus wide resources.

Unfortunately for us (and fortunately for the future gen-


erations) we are the pioneers in this endeavor. Soonthese positions will be an obvious requirement for anymodern statistics department, given the nature of thefield. But in the meantime, don’t give up the battle!

With that fiery flame I conclude my final newsletter ar-ticle as chairman of the statistical computing section —it has been a pleasure serving you.

Trevor HastieChair, Statistical Computing [email protected]

STATISTICAL COMPUTING

Student PaperCompetitionThe Statistical Computing Section of the ASA will spon-sor a Student Paper Session next year at the OrlandoJoint Statistical Meetings in 1995. The topic of the ses-sion will (naturally) be Statistical Computing. Threeto four students will be selected to participate in thissession, which will include a discussant nominated bythe selection committee. Fees associated with registra-tion, accommodation and travel to the conference willbe awarded to the participants in this Session.

Call for PapersStudents at all levels (undergraduate, Masters, andPh.D.) are encouraged to participate. To be eligible,an applicant must be a registered student in the fall of1994. The applicant must be the first author of the paper.

To be considered for selection in the session, studentsmust submit an abstract, a six page manuscript, a re-sume, and a letter of recommendation from a mentorfamiliar with their work. The manuscript should besingle-spaced in a 10 point font with one inch margins(this is consistent with ASA’s Proceedings guidelines.)In the case of joint authorship, the mentor should indi-cate what fraction of the contribution is attributable tothe applicant.

All application materials MUST BE RECEIVED byJanuary 10, 1995. They will be reviewed by the Sta-tistical Computing Section Student Paper CompetitionAward committee. The topic of the paper should bein the area of statistical computing, and might be origi-nal methodological research, some novel application, orany other suitable contribution (for example, a software

related project). Selection will be based on a variety ofcriteria at the discretion of the selection committee, andwill include novelty and significance of contribution,amongst others.

Award announcements will be made January 23, 1995.The selection committee’s decision will be final. Therewill be a discretionary cap of $1000 on any given award,but it is anticipated that this figure should be more thansufficient to cover the expenses.

Students not selected for inclusion in the Session maysubmit their abstract and a registration fee to ASA byFebruary 1 if they plan to attend the Joint Meetings.Those abstracts must be submitted following the ASAabstract submission instructions described in AMSTATNews. Students selected for inclusion in the sessionwill receive further information about abstract submis-sion and fee waivers from the award committee.

Inquires and materials should be emailed or mailed toeither one of the following:

Trevor Hastie, Statistics Department, Sequoia Hall,Stanford CA [email protected]

Daryl Pregibon, room 2C264, AT&T Bell Laboratories,600 Mountain Avenue Murray Hill, NJ [email protected]

All electronic submissions of papers should be inpostscript.

BUSINESS MEETING MINUTES

Statistical ComputingSectionMinutes of the Monday, August 15, 1994 Section Ex-ecutive Committee Meeting, Toronto, CANADA

Attendance

Trevor Hastie, ChairMary Ellen Bock, Chair-ElectJames L. Rosenberger, Newsletter EditorTom Devlin, Continuing Education CoordinatorDeborah Swayne, Secretary-TreasurerKaren Kafadar, Publications Liaison OfficerJohn Rice, Program Chair-ElectRon Thisted, Section Representative 1994-96Russell Lenth, Council of Sections Representative


Rick Becker, Graphics SectionRoy Welsch, Graphics SectionMike Meyer, Graphics Section and Newsletter EditorTerry Therneau, 1995 Secretary-Treasurer

Elected to new offices for 1995

Chair Elect: Sallie Keller-McNultyProgram Chair-Elect: Robert TibshiraniSecretary/Treasurer: Terry TherneauCouncil of Sections Reps: MaryAnn Hill (1995-97),Michael Meyer (1995-96)

Assuming new offices in 1995

Chair: Mary Ellen BockProgram Chair: John Rice

Previous minutes The minutes of the June, 1994 meet-ing, presented by Deborah Swayne, were approved.

Membership as of 8/4/94

Stat Computing Section alone: 761Stat Graphics Section alone: 171Joint Computing and Graphics: 2362Total Computing: 3123

Treasury as of 6/30/94

Section Cash $36,450.31Joint Projects Cash $8,253.29

Student Paper Competition Trevor handed outbrochures announcing the student paper competition.

Council of Sections Meeting Ron Thisted and RussellLenth reported on the Council of Sections meeting thathad just taken place.

ASA plans to hire a consultant on organizations to sug-gest changes in the structure of ASA, with the particulargoals in mind of improving outreach and member ser-vices.

There was also a lot of discussion about certification.The membership was recently polled on certificationissues. Approximately 5000 members responded, andthey were roughly equally divided on the subject—buteach section is not equally divided. The section onquality and productivity is the most strongly in favor ofcertification, and it is possible that the ASQC may goahead and certify statisticians whatever the ASA does.A resolution was passed and then tabled at the boardmeeting, saying that the ASA will have discussions withother societies about certification.

The Council of Sections representatives were asked tobring feedback from section meetings to a second meet-ing to take place toward the end of the joint meetings.We agreed to send back a statement saying that we en-

courage sections to work with other societies on theircertification issues, taking in effect a consulting role.

Continuing Education Tom Devlin reported that thesection’s continuing education courses were very suc-cessful at the Interface conference this year, and he’sinterested in some suggestions: At what other meetingsmight we sponsor courses? What new courses could weoffer? What new avenues could we use to find morespeakers and topics?

John Rice suggested a topic: non-linear random effectsmodels are of great interest in the pharmaceutical indus-try. Two avenues were mentioned as well: Could wesomehow make use of the world wide web? Shall wesolicit suggestions and volunteers in Amstat News?

CE: Cancelling courses Ron Thisted said that there hadjust been a discussion about continuing education at theCouncil of Sections meeting. Many short courses werecancelled by ASA last year because enrollment was low(below 25). This caused much unhappiness, and a deci-sion was just made that if enrollment falls below somethreshold, the sponsoring section can choose whetherto cancel the course or to split any financial losses withASA.

CE: Student scholarship program Tom continuedwith a proposal to extend the section’s support of stu-dents who attend our continuing education courses at ei-ther the Interface or at the joint meetings. We approveda proposal to pay for the conference registration, as wellas the course registration, for up to twenty students peryear.

CE: Additional support for course presenters Wenext discussed proposals for expanding financial sup-port for presenters of continuing education courses atthe joint meetings. Currently, speakers receive fromthe ASA a $500 honorarium and up to $1000 to covertravel and lodging costs. (This money is shared amongall presenters.) We approved a proposal to give someadditional money to presenters, to help them defraynecessary expenses: $250 for each speaker up to threespeakers. Despite describing the money as expensemoney and not an honorarium, we will not ask for re-ceipts. (Note that this funding applies only to the jointstatistical meetings and not, as yet, to the Interface con-ference.)

Both of these plans—the student scholarship programand the additional support for course presenters—willbe announced both in the newsletter and in AmstatNews.

Newsletter editors Two new newsletter editors shouldbe identified by the end of the year. It was also agreed


that some arrangement should be made so that the termsbegin to be staggered instead of beginning and endingin unison; for example, one of the new editors couldhave a two-year term and the other a three-year term.

Term of Secretary-Treasurer We clarified the previousdecision to extend the term of the secretary-treasurer totwo years in this way: A new secretary-treasurer willstill be elected every year, and each one will stick aroundfor two years, acting as Past Secretary-Treasurer duringthe second year. (Note: check by-laws)

Internet connection for the ASA Mike Meyer com-mented that electronic communication with ASA of-ficers is quite difficult, and volunteered to investigategetting them a better internet connection. He suggestedthat we consider paying for the better connection for ayear in order to encourage them to proceed. We agreedto consider it once he had investegated the issue.

Deborah SwayneSec/Treas, Statistical Computing SectionBell Communications [email protected]

BUSINESS MEETING MINUTES

Statistical GraphicsSectionMinutes of the Monday, August 15, 1994 Section Exec-utive Committee Meeting, Toronto, CANADA

Attendance

Rick Becker, Linda Clark, Bill DuMouchel, Jane Gen-tleman, Colin Goodall, Dave Hoaglin, Sallie Keller-McNulty Robin Lock, Mike Meyer, Barbara Milton,Sally Morton, Bob Newcomb, David Scott, Roy Welsch

Election Results

Roy reported the election results for 1995:Chair elect: Bill DuMouchelProgram chair elect: Steve EickSecretary/treasurer: Bob NewcombPublications Officer: Deborah DonnellCOS Representative: Colin Goodall

Committee Reports. Bill Tucker, Vice Chair of theCouncil of Sections gave a brief report on the functionof the COS, and urged members to interact with/givefeedback to our COS representatives.

Data Analysis Exposition. Robin Lock will be takingthis over for next year. As of yet there is no datasetfor next year’s competition. Since it is already late inthe year, we need to start publicizing this asking forsuggestions. Notes should go in AMSTAT News andthe computing/graphics newsletter. David Coleman hasdone this in the past.

Poster Competition. Jerry Moreno attended the meet-ing to give a report on the activities of this committee.He gave an oral report and also had a written version,which is available for further detail. Some issues thatwe discussed were:

Roy thanked Jerry on behalf of the section for all of hisefforts on this project. It has been, and continues to be,a great success.

Last year the executive committee approved John Schol-lenberger to replace Jerry, who wishes to step down.

Mike Meyer will include Jerry’s report in the nextnewsletter, along with pictures of this year’s winners.

After some discussion of the $3 entry fee, and whetherit was prohibitive to some, the executive committeeagreed to leave the fee as it is and generally give thecommittee freedom to do things on their own.

Committee to Nominate Fellows. Lorraine Denby be-comes the third member of this committee along withLarry Cox and Dan Carr. Andreas Buja, nominated bythis section, was made a Fellow this year.

Dynamic Graphics Video. Sally Morton reported onthe activity of the video lending library. There werethree new videos submitted this year. She is currentlyarranging for an Australian university to copy the videosinto PAL format. Some issues raised were:

She is looking for suggestions on making a historyvideo.

She’s also looking for suggestions on how to advertisethe video library.

Sally would like to step down from this committee inabout a year, so we need to find a replacement.

Continuing Education. Karen Kafadar has agreed toreplace Abbe Hertzig on this committee. We are look-ing for suggestions of courses the Section could spon-sor. One idea mentioned was to do something at SIG-GRAPH.

Charter. Dave Hoaglin reported on the proposedchanges to the charter of the Section, and had copiesavailable. The changes are modest and will be voted onby Section members next Spring.


Newsletter. Mike Meyer reported on the status of theComputing/Graphics Newsletter. A replacement forMike will need to be found by the end of next year. Theyhave made the last two years’ editions of the newsletteravailable on Statlib and the World Wide Web.

Treasurer’s Report. Linda Clark reported the follow-ing amounts in the Section accounts:

Graphics Section account $33,881.30Joint computing/graphics account $ 8,253.29

There was some discussion on whether we should con-tinue to solicit funds for the mixer from software ven-dors. Also, the questions was raised as to how to decidethe amount of money to budget for the mixer. Roy willlook into this.

Program reports. Sallie McNulty reported on two ten-tative invited sessions for 1995. She also said that weare at risk of losing sessions because of low attendanceat graphics section talks, and also because the numberof contributed papers has decreased. Sallie also com-mented on the need to have better equipment availablefor speakers at our sessions. Good graphics talks of-ten need more than just an overhead projector, and weshould look into the availability and cost of additionalequipment.

Report of the council of sections representatives.Sally Morton and Jane Gentleman reported on the COSmeeting held earlier in the day.

Here are some highlights:

Barbara Bailar has resigned effective Jan 15, 1995.

The idea of a second senior staff position being addedhas been postponed until further study by a consultant.The ASA board wants member feedback to Mitch Gailby early November and is circulating questions to guideinput.

Results of the survey on certification show opinions arevery split. The board put forth a motion(tabled untilDec) to discuss the issue with other societies (ASQC,IEEE).

Continuing education: course cancelation policy willdepend on 1) minimum enrollment or 2) the section tak-ing on the loss if enrollment is too low. Currently thesection makes $2,000 if there are 35 or more enrolled.We could subsidize enrollment by giving scholarships toencourage participation or buying slots to keep a courseform being canceled.

JCGS graphics paper prize committee. Roy will asksome people to serve on this committee.

Expenditures for outside speakers. There was some

discussion about what the policy should be on payingspeakers. Bill DuMouchel suggested we give the pro-gram chair money to 1) find outside speakers (whetherthey need to be ASA members would be up to the dis-cretion of the chair) 2) promote section activities. Theprogram chair will be informed that they can requestfunds for these purposes. The executive committee didvote to approve funds for this purpose.

New section activities. Some suggestions on how tospend section funds:

Mike Meyer - 1) offer a free one year membership tothe section 2) connection to the Internet for the ASA(possibly jointly with computing). Bill DuMouchel - 1)money to help support the poster competition 2) help onthe poster competition brochures

Linda ClarkSec/Treas, Statistical Graphics SectionAT&T Bell [email protected]

MIXER NOTES

Mixer at the JointMeetingsThe joint business meeting and mixer at the joint sta-tistical meetings of the ASA in Toronto, August 1994,followed the traditional pattern. A short reporting ses-sion from the section chairs, Trevor Hastie for StatisticalComputing and Roy Welsch for Statistical Graphics,brought to the members in attendance news from therespective executive meetings held earlier in the day.Following the reporting and announcements, the mixerbegan with a random drawing for door prizes, with thewinners receiving various books and materials donatedby the vendors at the meeting.

Following is a list of publishers and software companieswho donated door prizes for the mixer:

� Academic Press, Inc.� Addison-Wesley Publishing Co� Chapman & Hall� Duxbury Press� Irwin Publishing� John Wiley & Sons� Marcel Dekker, Inc� Oxford University Press


� Prentice-Hall and MacMillan� Springer-Verlag, Inc.� W.H. Freeman & Co� SAS Institute, Inc� BMDP Statistical Software, Inc

Software companies who made cash donations for themixer are following:

� BMDP Statistical Software, Inc.� SAS Institute� SPSS, Inc.� Stat. Graphics Corp� Systat, Inc.� Cytel Software Corp� Resampling Stats

We extend our sincere thanks to these companies forhelping to make the mixer a success!

Debby and LindaSec/Treas, Computing and Graphics

CONTINUING EDUCATON

Statistical Computingby Thomas F. DevlinStatistical Computing CE Chair

Short Course Proposals SoughtProposals for short courses on topics in statistical com-puting or statistical graphics are invited for presentationat the Interface ‘95 in Pittsburgh, on June 21. Thisyear’s themes include graphics, manufacturing and theenvironment. If you are interested in presenting a 2 houror 4 hour course, please send a proposal immediately,including topic and course outline, to the address below(email preferred).

You may also be interested in presenting a course atanother professional society meeting. Courses may beeither one-day (6 hours of instruction) or half-day (3hours) in length. They may be based on a new text,however that is not a requirement.

In addition to a modest honorarium and travel expenses,offering a course provides: professional recognition foryourself and the Section, an opportunity to share newideas with colleagues, exposure to a wide audience of

statisticians, a service to the profession and the Section,and an opportunity to promote statistical education andstatistical thinking.

Proposals, inquiries and suggestions should be sent to:

Thomas F. DevlinCE Chair, Statistical Computing SectionMathematics & Computer Science Dept.Montclair State UniversityU. Montclair, NJ 07043Voice: [email protected]

GRAPHICS SECTION RESOURCES

Graphics Video libraryIn the August 1993 issue of this newsletter (Vol 4, No.2, pp. 7–8) we published an article about the StatisticalGraphics video library, and listed the titles available.Here we present some updated information about thelibrary and the new titles. For the full list of titles werefer you to the earlier article.

Information about the library is available via e-mailfrom Sally [email protected].

New SubmissionsWe have had three new submissions to the library sinceAugust 1993:

Edge Information at Landmarks in Medical ImagesFredL. Bookstein and William D. K. Green, University ofMichigan Medical Center (1993). 26:15 minutes runtime.

Grand Tour and Projection Pursuit Dianne Cook, An-dreas Buja, Javier Cabrera, and Deborah F. Swayne,Bell Communications Research (1993). 19:00 min-utes run time.

Exploring Time Series Using Interactive Graphics An-drew McDougall and Dianne Cook, Rutgers Univer-sity (1994). 16:00 minutes run time.

The Dynamic Graphics Video Theater at the 1994 ASAMeetings ran from 8:30am to 6:34pm, and includedall library videos in chronological order. The tenminute ’Tokyo Data Map’ was shown continuously from6:04pm to 6:34pm to finish the Video Theater.


Borrowers and PAL Satellite Library BranchWe have approximately ten borrowers per year, mostlyfrom the U.S. with some interest from Korea. Atpresent, we are arranging for an Australian Universityto copy all the videos, or perhaps just those post-1987,into PAL format. The University is willing to act as oursatellite library branch, lending out the PAL videos toother PAL borrowers from Australia or New Zealandfor example.

New DirectionsThe Dynamics Graphics Video Committee welcomescomments from Computing and Graphics section mem-bers regarding the video library. If you have sugges-tions on how to improve the library, please send themto Sally [email protected].

Rules and Contents for the LibraryThe Video Lending Library of the ASA StatisticalGraphics Section.

The library is currently composed of the material thathas appeared in the ASA Dynamic Graphics Video The-ater, but over time will pick up new contributions. Thelibrary also serves as an archive where the material willbe preserved in a central and accessible place. Thelibrary contains 1/2 inch VHS video tapes and, for se-lected tapes, an abstract, related publications and tech-nical reports—even comments from previous viewers.

The video tapes are great for educational institutions:several people have used them as the basis for a “statday” production, for statistical computing classes, oreven for a video lunch complete with popcorn.

Lending policies for the library are:

� Loans up to 3 weeks are permitted� A limit of 3 videos at one time is permitted (with

exceptions for statistics day presentations wheremore videos will be available for a shorter period)

� The borrower will be responsible for includ-ing postage payment along with the request forvideos.

� Borrowers who are not members of the GraphicsSection will also be asked to pay a lending fee of$10; borrowing from the video library is one ofthe many services that members receive in returnfor their section dues.

Authors of the video tapes that are distributed as part ofthe lending library have given permission for the libraryto lend the tapes but have NOT authorized extra copy-ing. After viewing a tape, should you wish to acquire apersonal copy, you should contact the author.

To find out more about the video library, or if you haveyour own video tape that you think should be part of thevideo library, please contact Sally.

Sally MortonVideo Librarian, Statistical Graphics310-393-0411 x7360The Rand CorporationSally [email protected]

FEATURE ARTICLE (Cont.): : :

CONTINUED FROM PAGE 1

The WWW is a mechanism over the Internet by whichmultimedia information can be disseminated from siteto site. Through the use of client programs such asthe National Center for Supercomputing Application’s(NCSA) Mosaic and Netscape, information, includingdocuments, sounds, images, and video clips, can betransferred between machines world-wide. Informationis requested automatically from thousands of server ma-chines on the network through the use of the client pro-gram Mosaic. The server machine informs the clientwhat type of information is requested, i.e. whether it istext, pictures, video, hyper-text, or audio, and the clientdisplays or presents the information in the appropriateform using the appropriate viewer program, such as xdvifor dvi files, ghostview for postscript files, xv for gif,tiff, and jpeg graphics files, netaudio for audio files, etc.

Statistical data can also be viewed as a type of infor-mation, and therefore sent through the Internet and dis-played with a “viewer”. For example, one can retrievedata using FTP or Mosaic from sites such as Statlib, theCensus bureau, or the NIH. The natural way to “view”this information is with a statistical package! The pack-age provides the means of summarizing and displayingthe data for the user to understand.

Using the concept of duality, statistical procedures suchas linear regression or robust estimation can also beviewed as information—instead of the server supply-ing the data, imagine instead the server supplying theprocedures. Statistical procedures can thus be down-loaded and illustrated on data, as opposed to data beingdownloaded and viewed on the procedures. Interactivestatistical packages therefore become the tool which al-lows the data and procedures to interact and therefore


can be considered “viewer” programs for this informa-tion.

Teaching Paradigm

Interactive statistical packages and interesting datasetsprovide a rich combination for teaching and learningstatistics. Just as one interacts with a postscript or dviviewer to change pages, resize the view, and act on thevisualization of the document, statistical packages canbe thought of as a means of acting on the visualization(“view”) of the data and the elucidation of statisticalprocedures.

Hypertext documents are designed for user interaction.In the context of learning statistics, the student can viewdemos, cut and paste sample commands into the pro-gram, try out suggestions, quickly find the definitionsfor new words, and follow up on related concepts. Forexample, in HTML, help for concepts and keywords canbe linked inside the document, so that unlike a textbook,simply clicking on the word allows the user to easily finddefinitions for unfamiliar concepts. In addition, just asone might follow a link from the document to a def-inition, a picture, movie, or sound-byte, one can linkdata for an assignment as a similar piece of informa-tion. Instead of viewing a picture, the data is “viewed”or elucidated, using statistical procedures such as EDA,estimation, or hypothesis testing. The student can loadthe data and procedures as needed and will save or storeinformation only if it is useful. Changes to data sets andupdates to procedures and macros can be distributedthis way. An example of this is the WWW interface toStatlib.

This facility gives students online help for statistics pro-grams and related packages. Existing documentationis easily converted into HTML, the hyper-text markuplanguage. This is relatively trivial in the case of plainASCII text and LATEXdocuments for which an automatictranslation program exists. In addition, there is a wealthof documentation which has already been converted.

One motivation for this work stems from a need tosupply students with reference information about statis-tics and statistical computer packages. The three ap-plications which the first author is currently develop-ing are instructions and aids for statistical packages(Xlisp-Stat and Minitab), supplemental materials suchas course handouts or additional notes, and finally ad-ditional exercises which are optional and might assiststudents in performing the required exercises. Thesematerials have been designed on the assumption that thecourse has an accessible instructor, though these materi-als are especially suited to distance learning. An entire

course could be taught over the Internet—one modelis the Introduction to Object Oriented Programmingusing C++, available through the WWW , (connectto these materials using http://uu-gna.mit.edu:8001/uu-gna/text/cc/index.html).

Mean versus Median Example

Consider a tutorial on describing differences betweenthe mean and the median to describe the center of adistribution. A tutorial document discusses the meanand median for continuous data, which is not differentthan standard class notes or handouts. However, wecan embed hypertext links to separate documents whichfocus specifically on the mean and the median to helpremind the students of the definitions and properties ifnecessary. Exhibits can also be used to illustrate thepoint, just like a textbook. However, unlike a textbook,a demonstration program is linked in, which the studentcan use to explore how changes in a single data valueinfluences the sample mean and median. This demon-stration can use either text-based or graphical input forinteractively changing the values of data points one at atime.

Example on LineIn fact, this tutorial or lab session is available online. Totry it out, use a WWW client like Mosaic to connect tohttp://www.stat.psu.edu/statlearn/and pro-ceed to the center of distributions document. The sec-tion below provides more details for the new user.

Detailed Demonstration InstructionsTo try out the demonstration, three main steps must betaken.

1. Retrieve and Install an HTML viewer such asNCSA Mosaic.

2. Retrieve and Install Xlisp-Stat on your machine.3. Retrieve a copy of the modified mailcap and

Xlisp-Stat viewer script.

The first step has been documented in a number of otherplaces and so will not be repeated here (See e.g. Meyer,M.M. “Mosaic and WWW”, Statistical Computing andStatistical Graphics Newsletter, 4(3):19-20, 1993) Afterthe first step has been taken, connect to the documenthttp://www.stat.psu.edu/statlearn/ for fur-ther instructions for steps 2 and 3. The second step isalso documented well in Jan de Leeuw’s article on page16 of this Newsletter. Pointers for retrieving Xlisp-Stat source, binaries, the modified mailcap file, and theXlisp-Stat viewer script are provided. In addition, linksto connect the tutorial session on Mean versus the Me-dian are given.


Statistical Information as Objects for ViewingImplementing procedures for displaying statistical in-formation can take two approaches, referred to as server-side and client-side, depending on where the statisticalpackage resides.

For the server-side approach, the server computer doesall of the computations, and the client only needs to beable to view the resulting text, graphics, movies andsound that are generated from the client’s request. Thiscould require a complex user interface in order to handlethe general problem. Of course packages are availablewhich aid in the design of the HTML forms used forthis interface. The basic idea is that the forms usedas the interface for generating the statistical macro willproduce the desired results from the statistical package.The server machine handles the statistical computation,formats it as a document, and returns the document tothe user. This resulting document could include linksto other text, graphics, and perhaps audio or video clipresults which are pertinent replies to the statistical query.

The second approach, referred to as the client-side ap-proach, is to have the client computer run the statisticalprogram locally, and have the serving machine send anyspecial data and statistical procedures required to com-plete the exercise. This has the advantage of minimizingthe load on the server. However, it requires the client totrust the server, since the server will have some accessto the client’s machine; this is because most statisticalpackages allow one to escape to the shell (or MS-DOS,in the case of IBMs and compatibles), and there arepotential security problems. For example, one coulddownload a virus or a malicious shell script, and caremust be taken to guard against this!

The problems of the first method include a large loadplaced at a central site and the difficulties of construct-ing a general interface. Software licensing might alsobe a big problem! However, the second method is notwithout its faults, as it both produces potential securityproblems and needs technical details to be worked out sothat the statistical package does not have to be restartedfor each piece of information viewed. The latter is acurrent project.

ImplementationWe choose to use Xlisp-Stat as the basis for the client-side, and SAS for the server-side implementation.Xlisp-Stat is free and is available for most commonplatforms on the Internet such as Unix workstations,MS-Windows, and the Macintosh. In addition, thereare extensive help files which are available for convert-ing to HTML, and can be used to learn Xlisp-Stat !

Client Side

The HTML specifications allow one to specify file typesidentified by suffixes, using the MIME specification formulti-media mail. This is done at two points; the servermust acknowledge and inform the client of the type ofdocument being transfered (plaintext, HTML , postscript,etc) and the client must know what should be done withthe document. The server side description is done byspecifying a new type in the configuration file (differentfor different http servers) and the client-side configu-ration is done by specifying a viewer for “statisticalinformation” files through the system or user’s .mailcapfile (on a Unix system). The final necessary tool is ascript which runs Xlisp-Stat inside an xterm window (aterminal window in X11 which contains an interactiveshell, like a dumb terminal), to be the Statistical Viewer.

The user starts up the client and opens the server whichcontains the course notes or other information to bebrowsed. At this point, the procedure is similar to read-ing a book, except that the experience is easily madenon-linear.

Server Side

For the server-side mechanism, scripts must be writtento take the input from the forms interface and producea batch file suitable for the statistical package. Thenthe script must reformat the output from the statisticalpackage. The demonstration uses SAS, with an inter-face developed for the script language TCL. This isbased on work done at the Department of Medicine,Section for Artificial Organs, at the Hershey MedicineCollege at Penn State. These scripts were originally de-signed to construct an easy interface to SAS for residentswho were interested in graphically examining a body ofmedical experiment data. For the server-side construct,contact the first author for sample scripts written in TCLfor the SAS system.

DiscussionIn conclusion, the Internet and the World Wide Web canmake a fine supplement to teaching a course. Thereis much useful information out there, and a little timespent on collecting it can save instruction time.

Many people are starting to produce documents; the oneadvantage is that in general, others can use the materialfor their courses, so that people do not have to contin-ually re-invent the whole wheel. There is work on thiscurrently being done at UCLA and U Minnesota, as wellas both in the Penn State Statistics and Biostatistics de-partments, and probably others are developing materialsthat we are not familiar with.


The paradigm of being able to treat statistical proce-dures and data as viewable information is a useful one.One extension which is immediately possible based onthis work would be in the possibilities of multi-media e-mail. One could conceivably discuss a proposed analy-sis through e-mail, with the relevant information passedback and forth.

There are further mechanisms, not examined here,which can be extremely useful for distance learning andthe spread of knowledge. For example, there is a mech-anism, MBONE, similar to usenet news and WWW ,for transmitting audio and video (movies) over the In-ternet. For the course on C++ programming which wasmentioned, students communicated with the teachingassistants through email, “talk”, and the Internet RelayChat (the latter two are means for real-time communi-cation by typing). Other mechanisms include MUDsand MOOs (Multi-User Dungeons and MUDs, ObjectOriented), which are similar to the old text adventuregames, except that multiple players are involved, andobjects can be left by different users for each other. Forvisual communication, there is the CuSeeMe package,which transmits audio and visual between two worksta-tions, almost like a visual phone call. In this manner, avirtual classroom can be constructed.

Appendix: Technical Details for UNIX sys-temsIn order to use the server software developed at Penn State,the client must be able to translate the document. For thispurpose, a new information type must be defined for Mosaic,application/x-stat-xlisp, and the viewer needs to be specified.The viewer is called xlispstat.wrap to denote that it is a wrap-per around Xlisp-Stat , i.e. a program which makes surexlispstat starts up properly. With this installed, the followinginformation should be put into a global mailcap file or into auser’s personal mailcap file (i.e. .mailcap in the user’s homedirectory).

# Statistical informationapplication/x-stat-xlisp; xlispstat.wrap %s

A sample mailcap file and the script for the wrapper are avail-able via anonymous ftp from ftp.stat.psu.edu in thefile /pub/statlearn/xlispclient.shar.

The other side of the story is that the server must know thatfiles ending in .lsp are xlispstat files. This is usually de-clared in the configuration file of the http server (the programthat serves hypertext documents to clients). This requires thatthe configuration files be modified. An example of the codeto be added for the NCSA HTTPD server is as follows:

# Added for statistical information,# AJR (091694). EXPERIMENTAL!AddType application/x-stat-xlisp lsp

These lines must by placed into the srm.conf configurationfile. This would vary for different HTTPD servers.

One could do a similar thing for other statistical packages;while this transfers easily to Splus and Minitab, it also couldbe set up for SAS or other generally batch systems. For Splusor Minitab, it would suffice to choose a local suffix—perhaps.q for Splus and.mtb for Minitab. Then one could constructa universally accepted MIME type such as application/x-stat-splus and application/x-stat-mtb and write the respective in-terfaces.

Anthony J. RossiniCenter for Biostatistics and EpidemiologyHershey Medical Center at Penn [email protected]

James L. RosenbergerDepartment of StatisticsPenn State [email protected]

SECOND FEATURE

The Lisp-Stat StatisticalEnvironmentby Jan de Leeuw

A Question and an AnswerName a statistical package which

� is infinitely extensible, open, and comes with anelegant object-oriented extension language;

� is fast, use garbage collecting, and has a byte-compiler for additional speed;

� is interactive, dynamical, and graphical, and al-lows you to build to own graphical interfaces;

� is portable, and uses the native window systemon Mac, MSW, X11;

� is free.

After this blatant commercial beginning, everybody iswaiting for an answer. Here it is. Lisp-Stat !!

Some HistoryLisp-Stat is written in a combination of C and Lisp, justlike GNU Emacs. The basic language used to write theinterpreter and the build-in functions is C, and Lisp isused as the extension language of the system.


We shall not discuss the complicated history of the Lisplanguage. If you want to know more, you can read

˜/common/docs/Evolution-of-Lisp.ps

Note: In the sequel we shall refer to various docu-ments on our ftp server. The root for paths we give areftp://ftp.stat.ucla.edu/pub/lisp.

For our purposes it is sufficient to know that in theearly eighties Lisp systems were huge and expensive.Therefore David M. Betz decided in 1985 to write asmall and portable interpreter for a subset of the Lisplanguage. PC and Mac users could run their own Lispsystems by using his Xlisp, and they could experimentwith object-oriented programming as well. The inter-preter was written in C. Xlisp 2.0, released by Betzin 1988, has many additional functions and data typesbuild in, and the newer releases which are based on thework of Tom Almy, Niels Mayer, and Luke Tierneyare getting closer and closer to Common Lisp. Docu-mentation is in /xlisp/docs. There is a referencemanual (for Xlisp 2.1) by Tim Mikkelsen, and introduc-tion to Object-Oriented Programming (OOP) by TimMikkelsen, and a little manual on how to extend Xlispat compile time by adding additional C functions by JeffProthero.

In 1988 Luke Tierney published the first version of Lisp-Stat, which was documented in the technical report/xlisp/xlisp-stat/docs/xlispstatdoc.psTierney took Xlisp 2.0 as his starting point, and addedfunctions for linear algebra, optimization, statistics, andgraphics programming. Some of these functions areadded at the C level (i.e. at compile time) and some atthe Lisp level (i.e. at run time). Then in 1990 he pub-lished the book Lisp-Stat. An Object-Oriented Environ-ment for Statistical Computing and Dynamic Graphics.Unfortunately Xlisp 2.1g, which is used in the newestversions of XLisp-Stat, is so much larger than Xlisp 2.0that the account of the system given by Tierney in hisbook is quite incomplete. An update to the book is in/xlisp/xlisp-stat/docs/changes.ps.

The various names illustrate, by the way, the relation-ship between Lisp-Stat and Xlisp-Stat. Any Lisp sys-tem with the statistics and graphics added is a Lisp-Statsystem, but adding the Stat component to Xlisp givesXlisp-Stat. There is an (incomplete) version of Com-mon Lisp Stat, and we can easily image Elisp-Stat orScheme-Stat.

Some ComparisonsWe have called Lisp-Stat a statistical package. This issomewhat misleading, because there are so many dif-ferent types of packages, and Lisp-Stat is different from

almost all of them. It is not similar to the dinosaursSAS or SPSS, which apply relatively routine statisticsto huge data sets, with gigantic amounts of overhead. Itis also not similar to the many PC/Mac packages thatdo more or less the same thing as SAS and SPSS, onlyon a somewhat smaller scale, and with a little bit moregraphics. I think the only useful comparisons are withMatlab, Mathematica, DataDesk, and in particular withS-plus.

A fairly extensive comparison between Xlisp-Stat andS is contained in a collective review of List-Stat in theNovember 1991 issue of Statistical Science. The issuehas reviews by Baxter and Cameron, Weihs, Young, andLubinsky. Especially Lubinsky makes fairly detailedcomparisons with S. While S, especially in its S-plus im-plementation, is more complete and glossy, Xlisp-Stat isgenerally faster and has (much) better memory manage-ment. The speed advantage is more pronounced nowthan in 1991, because Xlisp-Stat has a byte-compiler inits newer versions, and because of its portability, it hasbeen rapidly set up on newer and faster systems. S hasmore statistics, Xlisp-Stat has more graphics. S is per-haps more appropriate for doing statistics, Xlisp-Statis more appropriate for producing statistics software,because it comes with a complete GUI builder. S isbetter for those who write papers in which statistics areapplied, Xlisp-Stat may be better for those developingstatistics.

Another important comparison is the language. Xlisp-Stat uses Lisp, a general purpose language with manyapplications. S uses the S-language, a special purposelanguage geared to statistics. I happen to think that ifyou work in a UNIX environment, then learning Lispis a good investment (think of Emacs, Maxima, and soon). In fact, some free Common Lisp systems (such ascmucl) are now available that can even do fast floatingpoint, and code for these systems is often very close toXlisp code.

What it comes with

Xlisp-Stat comes with everything in Xlisp-2.1g, includ-ing many of the functions in the various Lisp standards.In fact I would guess (but I am not sure) that the sys-tem contains well over 500 Lisp functions not relatedto statistics. This provides an extraordinary powerfulbasis for program development. Many of the func-tions are completely undocumented in the Tierney book,and even in the Mikkelsen reference manual. For in-stance, Xlisp now contains the system package fromCommon Lisp, with the many functions that come withit. Xlisp now can also save workspaces, similar to APL


workspaces, which makes it easy to customize and savework between sessions.

Another important addition to current versions of Xlisp-Stat is the byte compiler. The lisp code can be byte-compiled, and byte-compiled code generally loads andexecutes much faster than the original lisp. For codewith lots of loops and lots of consing, using the bytecompiler can mean speed ups of 500% or more.

Of course all the basic arithmetic and transcendentalfunctions are there, including the beta and the gamma.Descriptive statistics such a mean, median, standard-deviation, quartile, max, min, and fivenum are available.The numerical linear algebra functions are the usual,with inverse, determinant, solve, eigen, sweep, qr-decomposition, cholesky-decomposition. Arrays withunlimited dimensionality are native in Lisp, and thereare various primitives to handle indexing (this has beenextended greatly with add-on Lisp code, see below).There are two general purpose optimizers, one based onNewton’s method and one based on the Nelder-Meadsimplex method, and there is a somewhat idiosyncraticbut very useful (they tell me) module to approximatethe posterior mode using Laplace expansions.

A nice generalized linear modeling package is addedat the Lisp run-time level. It is similar to the nonlin-ear regression package because it inherits much of thefunctionality of the regression model prototype. Re-gression is one of the areas in which Lisp-Stat reallyshines. It forcefully demonstrates the elegance of anobject-oriented organization, and combination of theregression and graphical facilities has lead to manygreat tools. (Note: The OOP system in Xlisp-Stat isprototype-based, while the OOP in Xlisp is class-based.This means that actually two systems are available, butXlisp-Stat users will routinely use the prototype-basedone.) The recent Introduction to Regression Graph-ics by Cook and Weisberg is an impressive example ofthe power of these modules. It is a clear illustrationthat Xlisp-Stat is ideally suited to develop modules forstatistics teaching, modules that can be used by studentswho do not have to learn a single parenthesis of Lisp.

The uniqueness of Xlisp-Stat are its graphical interfacebuilding facilities. It will take some climbing of a rel-atively steep learning curve to use this effectively, butonce you are at the top (for somebody who has manyother things to do this will take at least a year, for othersit will take a couple of months and the usual amounts ofcoffee) an amazing tool chest is available.

There are also some weak points, which are partly un-avoidable and partly easy to repair. Interpreters runningtheir own shell are relatively closed systems, and conse-

quently communication with the outside world (input,output) is somewhat of a problem. Reading data intothe system is slow and tedious. This could be improveda great deal by moving some of the input facilities fromthe Lisp to the C level. There is no spreadsheet dataeditor, although it would be possible to add one. Thehelp system is rather primitive. And there are no facili-ties to make publication quality plots. Some of this hasbeen taken care of in the Lisp add-ons discussed below,but the basic facilities should really be part of the coresystem (which means part of the C code).

For many users it may be a nuisance that generalizedadditive models, cluster analysis, and many other ad-vanced techniques which are in S-plus or SAS are notin Xlisp-Stat. Again, this merely reflects the fact that thesystems have different objectives. SAS is for produc-tion, S-plus is for research and production, Xlisp-Statis for research and teaching. It also is a self-fulfillingprophecy. As the sections below show, extensions canbe easily added to Xlisp-Stat at various levels. C andFORTRAN subroutines can be loaded at runtime, andlinked dynamically into the system. This makes com-plete libraries of techniques available with a minimaleffort. But somebody has to make the effort. Tierneyis maintaining and upgrading the core system, which isa lot of work. Extensions have to come from the users.Once again I will mention GNU Emacs (or tcl/tk orX11) to illustrate that if a free and open tool is adoptedby a user community, while the core is maintained byexperts, we can create a tool which commercial compa-nies cannot possibly compete with.

Why I love it

In the first place because of the GUI building tools.Second, because of the Lisp language (but that’s justtaste). Third, because it is free. This may look trivialif you have money, but it means that we can give it toour students – not only to the graduate students, but alsoto the undergraduates. They can use it not only on our$ 10K elitist workstations, but also on their lowly 386or Performa machines at home. And fourth, it is open.We can see exactly what it does, because we have thesource. We can extend it at compile-time, at link-time,and at run-time level. If we get the latest greatest UNIXbox, or a Pentium, or a PowerMac, we recompile to usethe power of the new machine. It is easy, the systemis eminently portable. Because of these four reasons, Ipersonally don’t even think of S or S-plus as an alterna-tive.


How to get itThe complete Xlisp-Stat code distribution can be foundin

˜/xlisp/xlisp-stat/dist/

and binaries for various types of machines are in

˜/xlisp/xlisp-stat/systems/

There is also quite a bit of Xlisp-Stat code onstatlib in the directory xlispstat. (Note: Bythe way, ftp.stat.ucla.edu is no competitor ofstatlib, it is more specialized and may be more com-plete in this area.)

We try to keep a collection of everything con-nected with Xlisp-Stat in the ftp directories. Thereis always a version of the source code, andthere are binaries for selected systems. We ac-tually mirror the pub/xlispstat directory fromumnstat.stat.umn.edu. As the path to the direc-tory shows, there are other lisp systems besides xlisp,and other xlisp extensions besides xlisp-stat. Our servercan also be reached viagopher.stat.ucla.edu orwww.stat.ucla.edu.

CodeThe most interesting sub-directories on our server are

˜/xlisp/xlisp-stat/code/contrib/

and

˜/xlisp/xlisp-stat/code/homegrown/.

In homegrown we have Xlisp-Stat code written atUCLA. This includes:

� the SIR package by Ker-Chau Li and his students,for Slicing Inverse Regression;

� the Terrace package by James Hilden-Minton forHierarchical Linear Modeling with diagnostics;

� the Gifi-gardens package by James Hilden-Minton for MVA with optimal scaling;

� many utilities by Jan de Leeuw for monotone re-gression, b-splines, the basic APL array opera-tors, multidimensional scaling, nonlinear MVA,FORTRAN-type formatted input, log-linear mod-eling;

� stem-and-leaf-plots by David A. Betz;� bootstrap/jackknife algorithms by Jan de Leeuw.

This includes bootstrap-t, percentile, abc, bcaconfidence intervals, bootstrap and jackknifestandard error, bootstrap and jackknife bias es-timation, bootstrap prediction error estimation.

The bootstrap algorithms for regression are writ-ten as methods for regression-model-proto. Thereis also a graphical bootstrap demo. And there iscode for permutation and bootstrap hypothesistesting. There are also versions of Randy Sit-ter’s algorithms for bootstrapping in survey situa-tions with clustered and stratified random samples(without replacement).

� Didactic Modules by Jan de Leeuw and JasonBond. Moving the regression line, moving pointsin regression scatterplots, the central limit theo-rem.

� UCLAData. Evolving package for IntroductoryData Analysis teaching by Jan de Leeuw.

� Help. Evolving graphical help system for Xlispprogrammers and Users by Jan de Leeuw and Ja-son Bond.

In contrib we have xlisp-stat code written by others.This includes everything on statlib, and then some. Wemention

� axis: Introductory statistics package by RobertStine (Wharton).

� bpois: Baysian Poisson Regression using theGibbs Sampler by Hani Doss (OSU) and B.Narasimhan (PSU).

� cw: Dynamic Graphics for Regression, by Den-nis Cook and Sandy Weisberg ( Minn). Calleddyndiag on statlib. This has been superseded bythe much-improved code that comes with theirrecent book.

� eltoy: Introductory Statistics and Bayesian Elici-tation, by Russell Almond (U Wash).

� expsurv: Exploratory Survival Analysis by K.Neely Atkinson (U Texas).

� hasse: F tests and ANOVA tables for factorialdesigns by Philip Iversen (Iowa State).

� jflisp: Three packages for Model Selection inRegression Analysis by Julian Faraway (Univ.Michigan).

� naras: Code contributed by B. Narasimhan (seealso bpois). This includes utility code for slidersand boxplots, as well as packages to demonstrateMarkov chains, and to fit kernel density estimates.There is also a package for programmers whowant to dynamically link arbitrary C functions atruntime.

� plottools: Don’t know who wrote it. It providesnice graphical windows to selects symbols andcolors.


� puranen: Some demos by Juha Puranen (UnivHelsinki). Demoes Box-Cox and some of thecommon densities.

� simulation. General purpose simulation tool byRobert Stine (Wharton).

� students: Projects by students in Nancy Reid’scourse at the U Toronto. Code for repeated mea-sures with missing data, teaching demos, corre-spondence analysis, Weibull MLE, the generallinear hypothesis.

� surv: Code for survival analysis by E. NeelyAtkinson and Meg Gelder.

� times: Digital filtering code by Bill Hatch (Cole-man Research Corp).

� udina: gnuplot interface to produce beautifulplots from Xlisp-Stat by Jan de Leeuw and Fred-eric Udina. kde kernel density estimation systemby Frederic Udina.

� xlisp-to-s: Interface from Xlisp-Stat to somethingcalled S or S-plus or something. By Steve McK-inney (U Wash)

� young. Vista: Visual Statistics System by ForrestYoung (UNC).

Some of these packages are just snippets, some are ma-jor research efforts. Some have been abandoned inmid-stream, others are still being developed. I knowof quite a few projects that are going on that we canhopefully add in the near future.

If you have code to contribute, no matter howincomplete, imperfect, tentative, please send it [email protected].

ConclusionAny knowledgeable instructor can make Xlisp-Stat intothe ideal tool for teaching statistics at the graduate level.With a considerable investment from many persons, wecan make it into the ideal toolkit for developing intro-ductory statistics teaching modules. With another suchinvestment, we can make it the ideal tool for statisticsresearch (and small-scale production). It is imperativethat more statisticians start using Xlisp-Stat in the nextfew years, because otherwise it will either wither anddie or go commercial and freeze. This would be a majorloss for the statistics landscape.

Jan de LeeuwStatistics [email protected]

DEPARTMENTAL COMPUTING

Establishing a Serverby Michael Conlon

Should your department have a “server” — a computerthat performs some collection of tasks for other com-puters in the department? What will be gained? Howdifficult is a server to maintain? If you already havea server are you getting the most out of it? Do youoffer the right mix of services to your “clients” — thecomputers that receive services?

Benefits of a serverAs networked desktop computers become more com-mon, the role of the server continues to grow. A server(regardless of platform—Mac, PC or UNIX) can be ex-pected to offer one or more of the following services:

Print sharing

In some configurations, certain computers act as “printservers.” Other computers rely on these machines forprinting across the network. The benefit is reduced in-vestment in printers, and a potential to support multipletypes of printing — laser printers, high volume print-ers, large-format printers, color printers. To get accessto a variety of print devices, the client computers areconnected to a network and one or more print servercomputers connect to print devices. It is also possi-ble for many types of printers to be put directly on thenetwork without a print server computer.

Backup

A server typically is maintained by a system administra-tor. That person’s job functions should include backingup the disk space on the server. The server may also re-motely backup file space on client machines across thenetwork. This is a tremendous improvement over the“everyone for themselves” model of backup in whichsome people do a better job than others. Intelligentbackup systems pool hardware resources by only hav-ing tape drives on the servers and only backing up filesthat have changed (incremental backups).

File sharing

Disk space may be available on the server for individu-als to use. This space may be accessible by other users,using logon and permission systems to enable users tocontrol who they share files with. Such arrangementsare a great improvement over file sharing by file copying


— “can I get a copy of x?” and then when x changes,you don’t get the updated copy. In file sharing sys-tems, you are getting access to the actual file. As itchanges, you get the newest material. File sharing isnaturally important in collaborative consulting environ-ments where more than one person may be involved ina data analysis.

Software maintenance

Many servers simply provide software to client comput-ers. This is a variation on the file sharing idea above,but happens at the system level. By installing softwareon the server, the software is centrally maintained. Onlyone copy needs to be updated. All client machines re-ceive the software at run-time — no copying – so thateach client is always running the current version and allclients are running the same version. This scheme alsomakes accounting for software much easier than dis-tributing the software out to individual hard disks. Andpotentially large amounts of disk space can be saved.The software collection at the Department of Statisticsat the University of Florida currently occupies approx-imately 2 gigabytes of disk space. We have 75 clientmachines. Even if it were feasible to put all that soft-ware on each machine (it isn’t), we save 150 gigabytesof space by managing it centrally.

Internet services

Servers are increasingly important to provide basic net-work services such as email, access to Usenet news-groups, and for serving web pages. While some of theseservices may be available from corporate or Univer-sity mainframes, the services available from these otherproviders tend to be inferior to what can be providedlocally. Few mainframe email systems support MIME— the Internet standard for multi-media mail, whilemany department mail systems do support MIME. Simi-larly, establishing web services on a department server isstraightforward, whereas interacting with a mainframecomputer for web services may be quite difficult.

Server modelsRegardless of platform, the organization of client com-puters with respect to servers can generally be classifiedinto three “models.”

No server

Here every client operates independently, as comput-ers did before networks. A local printer, a local tapedrive, no file sharing, no large software library, Inter-net services either provided on the desktop or not atall. Many people continue to operate happily this way.When considering the establishment of a server, one

needs to think clearly about the impact on such peopleand whether they will benefit from access to a server. Inmy college, many of the PCs operate this way. They areon the net with telnet and ftp and may have mail servicefrom a mail server, but all other services are providedlocally.

Weak server

The server provides some services but not all. The keyfeature of this model is that the desktop computers con-tinue to operate normally if the server is not available.Software is installed locally rather than accessed fromthe server. If the server is unavailable, clients may notget email, but they will still be able to access data (nofile sharing) and perform analyses (with local software.Of course some services may be available (access data)and some may not (printing) depending on the mix ofservices offered by the server. This model makes sensein places where the network is unstable — if the serveris unavailable because the network is down, people con-tinue to work. Software must be installed locally, so thevariety of software is limited. Many of the Macintoshcomputers in my college operate in this model.

Strong server

Most services come from the server. If the network orserver is down, the client machines can not be used.Sounds risky, doesn’t it? Actually, our departmentservers (three) have a better uptime record than thecampus mainframe (which is considered rock-solid).Network outage is a more serious problem. We havedesigned our system so that we have a server in eachmajor location, so that if the interconnecting networkis unavailable, each work group continues to function.Almost all of the UNIX computers in my college oper-ate in this model. PCs on Novell Netware often operatein this model, receiving necessary application softwarefrom their server.

The downsidePrint and file sharing, Internet services, backup and ac-cess to a large centrally maintained software collectionare attractive reasons to have a server. There are twomajor problems.

The first is cost. A capital outlay for server softwareand hardware and network connections to the clientsmay be be required. A continuing outlay for mainte-nance contracts, software renewals will be needed. Themost important cost is the time and effort of a competentsystem administrator to oversee the server and its parts.This is, by far, the largest part of the investment and thelargest deterrent to the second risk.


The second risk is server failure. If a strong servermodel is established so that client systems will not func-tion without the server, and the server is not available,department members will sit around unable to use thesystem — just as we did twenty years ago when themainframe was down. Many of us have memories ofthose days and no desire to return to them. The risk ofserver downtime is substantial. A good system admin-istrator is the best protection against server downtime,since most server problems are related to software andits configuration —- not to hardware failure. One canbuy a hardware maintenance contract and/or have spareparts available to recover from server hardware failure.Recovering from poorly configured server software isfar more difficult.

Multi-platform issues

In any consideration of services being offered centrallyfrom a department facility, thought should go into thevariety of machines in the department. Will everyonebenefit from a department mail server? If the server isa POP (Post Office Protocol) server, client software canbe obtained for Mac, PC and UNIX machines, enablingall to take advantage of the new department service.Proprietary vendor email systems rarely cover all plat-forms.

On balance

The benefits of server installation outweigh the riskswhere qualified system administration can be dedicatedto the task of maintaining the server. Strong servermodels are common because of the productivity advan-tages they deliver to their departments. Where systemadministration is weak or non-existent, the benefits ofdepartment servers are rarely realized.

Michael ConlonDepartment of StatisticsBox 100212 HSCUniversity of FloridaGainesville, FL 32610–[email protected] page:http://www.clas.ufl.edu/˜mconlon

TOPICS IN SCIENTIFIC VISUALIZATION

A Colorful Variation OnBox Plotsby Daniel B. Carr

Introduction and Background

The box plot provides a useful caricature for a batch ofdata. The word caricature is appropriate since the boxplot hides some distributional features, such as gaps,while calling attention to others, such as the median andquartiles. Since box plots receive such widespread use,it is natural for statistical graphics designers to see ifthey can make even minor improvements on the graph-ical representation. This article describes my efforts toimprove on the box plot. The article builds on the lastnewsletter article about grids (Carr 1994a). Both arti-cles come from a technical report on converting tablesinto row-labeled plots. Interested readers can find moredetails and coverage of other topics such as representa-tion of confidence bounds in Carr (1994b).

In terms of a brief history, Tukey (1977) introducedthe box plot. McGill, Tukey, and Larsen (1978)added the idea of notches for comparing medians.Frigge, Hoaglin, and Iglewicz (1989) described differ-ent choices for the five- number summary portion of thebox plot. Tufte (1983) provided alternative representa-tions that reduced the amount of redundant ink in thebox plot. His example suggested that others also searchfor a better graphical design. Tukey (1989) took up thechallenge to develop a better representation. His designstrategy used more easily distinguished graphical ele-ments for different components of the box plot. In onevariation an area symbol, the filled dot, represented themedian. This area symbol stood out from other graphicselements and helped when comparing medians. Tukey(1993) provided more variations on letter displays andbox plots and more thoughts on graphical comparisonof several linked symbols. As discussants of Tukey’sarticle, Becker and Cleveland (1993) reinforced the useof filled dots to represent medians and brought out fur-ther design considerations, such as sorting box plots bytheir medians. A simple look at the Becker and Cleve-land example shows that filled dots and sorting providea better basis for comparison. This history on black andwhite methods, the availability of color and variationsthat appear in software suggest that perhaps there is stillroom for improvement.


Design Objectives For The Box PlotA useful starting point in graphical design is to reviewthe design objectives. The box plot symbol is a compos-ite symbol that represents five summary numbers (me-dian, quartiles and adjacent values) plus outliers. Thedesign of the box plot involves selecting a compositesymbol to represent these five numbers. One design ob-jective is for viewers to perceive the box plot as a singleperceptual unit, as one symbol rather than five. A sec-ond objective is to ease comparison of elements withinthe box plot. For example, are the quartiles symmetricabout the median? Other objectives involve compar-isons among box plots: comparison of medians withother medians, upper quartiles with upper quartiles andso on. The multiple objectives complicate settling on anoptimal design. Design efforts here focus on promotingcomparisons among box plots.

Linkages that tightly bind elemental symbols into per-ceptual units make the elemental symbols difficult todisaggregate for comparison across perceptual units.Kosslyn (1994) discusses several principles of percep-tual organization of which two, proximity and good con-tinuation, are particularly relevant to composite symbolbinding. For example, the box in the box plot binds thequartiles together. The visual flow along the box lengthand along the lines from the box to the adjacent valueshelps people perceive the outliers as part of the samedistribution. A related principle for perceptual group-ing is visual enclosure. Enclosing the median symbolin a box helps make the symbol become an integral partof the composite symbol. Weakening the tight bindingprovides a means of encouraging comparisons amongbox plots.

One design objective is for viewers toperceive the box plot as a single percep-tual unit, as one symbol rather than five.

The filled dot suggested by Tukey (1989, 1993) is anarea symbol that creates an alternative perceptual group-ing:filled area symbols versus line symbols. The alter-native grouping counteracts the tight binding in earlyversions of the box plot. This approach suggests thenext step, using color to create alternative perceptualgroups. When color is available, using an area symbolfor the median becomes less advantageous. The boxplot design shown in Figure 1 uses black vertical linesto call out medians. The design represents the rangesbetween the quartiles and adjacent values using filledboxes. Different colors distinguish the median, valuesbelow the median and values above the median. A sim-ple white line shows the median comparison interval.The availability of color provides such options.

Several considerations motivate the particular designchoices. Comparing medians across box plots is oneof the most important comparison tasks. The currentdesign gives the medians a different color and shapefrom the other box plot elements. The distinctive sym-bol enables people to rapidly find the medians usingpreattentive vision (see Julesz 1986). One can focus at-tention on all the black line segments just as one couldfocus attention on all filled dots. In contrast to theearly box plot design, the line for the median comprisesthe vertical extremes of the composite symbol. Thisavoids total entrapment of the median symbol withinthe composite symbol. The vertical line segment ori-entation encourages the vertical visual flow needed tomake comparisons. Note that by itself the horizontalorientation of the composite symbol impedes verticalvisual flow. When comparing medians from adjacentrows the separation of the nearest endpoints is nearlyhorizontal. This helps in judging the correct distance.In the dot design it is difficult to judge the distance be-tween dot centers. When adjacent medians are nearlyidentical in value, they can be resolved with vernierresolution. Thus the symbol for the median facilitatescomparison.

As for the thicker box between the quartiles than be-tween the adjacent values, Tukey (1993) points out thatsymbols near the edges of the plot have higher visualemphasis. Stressing symbols in the center of the distri-bution provides balance. Turning the lines from quar-tiles to adjacent values into thin boxes is a small changethat accommodates the drawing of median comparisonlines as described below.

Distinguishing the left and right sides of the box plotwith color makes it easier to make comparisons amongthe respective halves of the distributions. For tasks likecomparing third quartiles, one restricts attention to lightred symbols. The almost instantaneous sorting by colornarrows the field for comparisons. Evaluation of theinterquartile range for a single box plot remains straightforward since the vertical line for the median and thecolor change does not seriously interrupt the visual flowalong the thick box between the quartiles. A few consid-erations should govern the color selection. The colorsshould be easy to distinguish from each other and themedian color. Colors for low and upper values shouldbe brightness matched (see Carr 1994c) and not fullysaturated (see Tufte 1990). When gray is the only op-tion, the most important objective is make the median adistinctive color. This suggests dropping the color dis-tinction between upper and lower values in gray levelplots.


EyeFemaleMale

Other EndocrineFemaleMale

NasopharynxFemaleMale

Salivary GlandFemaleMale

Nose, Nasal Cav.FemaleMale

Connective tissueFemaleMale

LipFemaleMale

Thyroid GlandFemaleMale

TestisFemaleMale

MelanomaFemaleMale

BoneFemaleMale

Multiple MyelomaFemaleMale

Other SkinFemaleMale

LarynxFemaleMale

Hodgkin’s DiseaseFemaleMale

EsophagusFemaleMale

KidneyFemaleMale

BrainFemaleMale

Oral CavityFemaleMale

Non-H.LymphomaFemaleMale

BladderFemaleMale

0 2 4 6 8 10

••••

•••

•••

•••••

•••

••

•••

••

••

••

•

•••••• •

•••

••••••

•••

•••

••

•••• ••

•

•

1950 - 1959

0 2 4 6 8 10

••••

•••••

•••

•••••••••

••

•••

•••••

••

•

•••

••

•••••••

••••

•• ••

•

••

••

•• •

•

•

1970 - 1979

Cancer Death RatesLow-Rate Sites By Sex For Two Decades

Box Plots Based On State Rates

White Line - 95% Test For Equal MediansSite

Sex

Age-Adjusted Rate Per 100,000

Figure 1: A Three-Factor Plot Showing Box Plots


The treatment of outliers in Figure 1 is standard. Analternative shows outliers as vertical line segments withthe same height as the adjacent value box. This re-duces overplotting, encourages vertical visual flow andplots symbols at their location centers rather than aroundthem. Open circles help increase visibility when thesegments are very short.

The proposed box plot design needs tobe tested in cognitive experiments to seeif it helps people to compare box plotsmore quickly and accurately.

Figure 1 uses a white horizontal line at the vertical centerof each box to represent the median comparison interval.This line representation works when a confidence boundgoes past a quartile. Notches appear strange when theyextend past quartiles. The internal line representationhas merits in addition to handling asymmetric distribu-tions. Since the confidence limits are symmetric aboutthe median, the confidence line allows more accuratecomparison of symmetry for the first and third quartiles.To assess quartile symmetry, one looks at the differencebetween the ends of the white line and the ends of thequartile box. According to Weber’s law, people cancompare these small differences more accurately (on anabsolute scale) than they can compare the larger dis-tances to the median. The symmetry of the white lineabout the median encourages mentally folding the boxplot about the median to judge symmetry. Since thewhite line is trapped inside the composite symbol, itdetracts little when comparing symbol outline elementssuch as third quartiles. With median comparison linesinside the boxes, comparison of lines may not be asprecise or easy as with other possible designs. This is areasonable compromise because approximate hypothe-sis tests are not the primary box plot feature. In otherwords the comparison lines are easy to ignore but serveseveral comparison purposes when needed.

The interpretation of median comparison lines is thesame as for notched box plots. The medians differ at thegiven significance level when the lines do not overlapand adjustments should be made in multiple comparisonsituations.

The Data Set, Plot Construction and Interpre-tationThe data set includes age-adjusted cancer rates for fourfactors: states (50), body sites (34), sexes, and decades(3). Figure 1 represents states using box plots, bodysites and sex using indented rows and decades usingpanels. The 68 body site by sex combinations stretchthe limits of the legible labeling and symbol plotting.

Consequently this plot shows values for the 21 bodysites with the lowest rates. Sorting by the male medianvalues for the 1950’s determines the body site order.The range of death rates remains large for these sitesand a log scale would be reasonable for the x-axis. Fig-ure 1 does not use the log scale in order to show thesymmetry of the confidence lines and the performanceof some box plot symbols under restricted resolutionconditions. Figure 1 shows just two decades to providemore resolution for some symbols. As it stands the onepage plot represents over forty percent of the 10,200values in the data set.

The interpretation is straight forward even in the over-plotting cases. For cancer of the eye, the median linescompletely cover the quartiles and the median com-parison lines. For males and females one can see thehorizontal difference between the median lines. This in-dicates that the hidden comparison lines do not overlap.The conclusion is that rates for males and females aresignificantly different. For the twenty comparable bodysites in the plot, male rates are higher with one excep-tion, the thyroid gland. The plot design makes it easy tocompare the rates for sexes. Comparing body site ratesfor a given decade and sex is not much harder since allvalues have the same scale. Comparisons across pan-els is a little harder but here the grid lines come intoplay (see Cleveland 1993 and Carr 1994a). Comparingbox plot elements against the grid lines makes the rateincreases from the fifties to seventies quite apparent.

Figure 1 provides a tremendous amount of informationbut is not a stopping point. To go further, one shouldstudy the national rates directly. The box plot outliersrepresent atypical state rates, so are of some interest.However careful comparison adjusts for estimate vari-ability that includes the influence of differing populationsizes.

Closing Comments

The plot that cannot be improved is rare. At thesame time attempted improvement do not always provesuccessful. The proposed box plot design needs tobe tested in cognitive experiments to see if it helpspeople to compare box plots more quickly and accu-rately. In the mean time, readers who like the de-sign should give it a try. S-Plus users can modifythe template for Figure 1. The cancer data, script file,rowplot function and box plot function are availableby anonymous ftp to galaxy.gmu.edu. The directory/pub/submissions/rowplot contains this andmany other row-labeled plot examples.


References

Becker, R.A. and W.S. Cleveland (1993), "Discussion ofGraphic Comparison of Several Linked Aspects: Alterna-tive and Suggested Principles," Journal of Computational andGraphical Statistics, Vol. 2, No. 1, 41-48.

Carr, D.B. (1994a), "Using Gray in Plots", Statistical Com-puting and Statistical Graphics Newsletter, Vol 5, No. 2, pp11-14.

Carr, D. B. (1994b), "Converting Tables into Plots." Tech-nical Report No. 101, Center for Computational Statistics,George Mason University, Fairfax, VA.

Carr, D. B. (1994c), "Color Perception, the Importance ofGray and Residuals, on a Choropleth Map," Statistical Com-puting and Statistical Graphics Newsletter, Vol 5, No. 1, pp16-20.

Frigge, M., Hoaglin, D. C. and Iglewicz, B. (1989), "SomeImplementations of the Box Plot," The American Statistician,43, 50-54.

Julesz, B., (1986), "Texton Gradients: The Texton TheoryRevisited. Biological Cybernetics, No. 54, 245-251.

Kosslyn, S. M. (1994), Elements of Graphic Design, NewYork, NY: W. H. Freeman and Company.

McGill, R., Tukey, J. W., and Larsen, W. A. (1978), "Varia-tions of Box Plots," The American Statistician, 32, 12-16.

Tukey, J. W. (1977), Exploratory Data Analysis, ReadingMA: Addison- Wesley.

Tukey, J. W. (1989), "Data Based Graphics: Visual Displayin the Decades to Come," Proceedings of the American Sta-tistical Association Sesquicentennial Invited Paper Session,Alexandria, VA: The American Statistical Association, 366-381.

Tukey, J. W. (1993), "Graphic Comparisons of Several LinkedAspects: Alternative and Suggested Principles," Journal ofComputational and Graphical Statistics, Vol. 2, No. 1, 1-33.

Tufte, E. R. (1983), The Visual Display of Quantitative Data,Cheshire, CT: Graphics Press.

Tufte, E. R. (1990), Envisioning Information, Cheshire, CT:Graphics Press.

Daniel B. CarrGeorge Mason [email protected]

NET SNOOPING

WWW viewers—Is thereone for you?by Mike Meyer

The Netscape BrowserIn the last issue of the Newsletter I hinted at a soft-ware offering from Mosaic Communications. Well,the infobahn is racing ahead at full speed. The com-pany changed its name to Netscape Communications(but kept the mcom.com Internet domain), releasedat least 3 beta versions of its software and won thepraise or wrath of just about every denizen of theInternet. Even Time Magazine, available online athttp://www.timeinc.com, in its December 51994 edition, has jumped into the fray, naming one ofthe Netscape developers, Marc Andreessen, to its list of50 people with the “requisite ambition, vision and com-munity spirit to help guide us in the new millennium”.At least the electronic version of Time, from which Icut and pasted the above quote, mentions Andreessen.I haven’t checked the printed version.

So what is all the noise about? Clearly not just a partic-ular piece of software—but I will limit my comments tothat.

Netscape is currently available atftp:ftp.mcom.comand several mirror sites. There are versions for mostpopular Unix platforms (using X windows) and Macin-tosh and Windows versions. The user interface is veryslick and remarkably similar across all three platforms,winning praise from many people. I regularly use theMac and Unix versions and I have no difficulty switch-ing between the two. The original Mosaic software,which by most measures would still be considered in itsinfancy, looks old and weary in comparison. Many ofthe people who praise the user interface also decry thememory requirements of the software. The Mac ver-sion requires 2Mb of memory and works best with 4Mb.Lots of people are spending lots of money on memory.Of course the people who complain about Netscape’smemory requirements probably have not tried to installthe latest commercial software from companies likeMicrosoft.

Netscape tries to do it all, and it largelysucceeds.

Why should Netscape require so much memory? Onereason is that Netscape tries to do it all, and it largelysucceeds. Where other WWW viewers rely on externalhelper applications, especially to show non-inlined gifs


and other images, Netscape includes its own integratedviewer. One effect is faster overall operation, and thememory requirements of Netscape are probably on a parwith another viewer plus helper applications.

Another area where Netscape receives both praise anddamnation is the issue of licensing. The current versionis (more or less) freely available to anyone. That seemslike a very generous offer. Furthermore the Netscapedevelopers have repeatedly stated on Internet bulletinboards that they intend to provide the browser for freeto not for profit organizations, like most universities.Even the quoted prices for Netscape Communications’http daemon look very reasonable. So, what is the prob-lem. Many people, including me, can not (or at leastcould not) find this promise as part of the Netscape li-cense or elsewhere in the Netscape documentation, andthat causes us to pause. Should we start relying on thissoftware for critical applications? An even more bizarreproblems arises for certain foreign nationals. Presum-ably because of US export restrictions on encryptionsoftware, Netscape is not allowed to be exported to cer-tain countries and even certain foreign nationals livingthe US are not allowed to use Netscape. At least that isthe how one of my Iranian colleagues, who is a profes-sional online information developer, reads the license.That is making it increasingly more difficult for him todo his job.

Where other WWW viewers rely on ex-ternal helper applications, especially toshow non-inlined gifs and other images,Netscape includes its own integratedviewer.

An area of Netscape that seems to have won largelypraise, is its performance over lower speed com-munications lines. One innovation is the handlingof “interlaced” images where one can quickly geta fuzzy picture that progressively becomes cleareras more of it is downloaded. Even StatLib(http://lib.stat.cmu.edu/), now uses inter-laced images. Quite fun. Another useful feature is theability to click on links even before all the documentis loaded. Savvy network providers are now realizingthat it is important to provide the important informationearly in the document, so that power Web cruisers cancontinue exploring.

Much more emotion is vented at the HTML (the na-tive authoring language of the WWW) extensions thatNetscape supports, and thereby promulgates. Somepeople like, and are actively using, the ability tohave more control over font sizes, and a number ofpage layout capabilities like centering text. Only

Netscape recognizes this new formatting—all otherHTML-compliant readers are supposed to ignore it. Inpractice it seems that the HTML extensions are at worstharmless to users who do not use Netscape, and quiteuseful to those who do. So, again, what is the fuss about?These HTML extensions are not the same as those pro-posed by other groups, and some people fear that themomentum behind the WWW will begin to dissipate ifmore and more incompatible HTML extensions prolif-erate. There is some merit to that argument. However,if enough people find the Netscape extensions to be use-ful I feel confident the extensions will be incorporatedinto the appropriate standards.

An area of Netscape that seems to havewon largely praise, is its performanceover lower speed communications lines.

Netscape is also the first client (and server) on the blockto offer secure communications—in this case by incor-porating encryption technology from RSA. This opensthe possibility of relatively secure banking and elec-tronic funds transfer. It may not be too long before youcan feel secure about entering your credit card into, say,Netscape and having the appropriate server contact thecredit card company, get approval for the transaction,bill you, and then ship you the information or productthat you ordered. The full text of the EncyclopediaBritannica (http://www.eb.com) is already avail-able, provided your educational institution has paid theappropriate fee. With Netscape it becomes feasible tocharge individuals for each transaction.

The final area where Netscape is either a winner, ora loser, is its ability to have multiple connections ac-tive at the same time. Many WWW browsers canhave multiple browse windows open, but so far onlyNetscape allows each of those windows to be retriev-ing different information at the same time. This can bevery useful. While I am waiting for an image from theLouvre (http://sunsite.unc.edu/louvre orhttp://mistral.enst.fr/louvre) I can alsoask Lycos (http://lycos.cs.cmu.edu/) tosearch through its database of almost 1 million uniqueURLs. Some people predict the demise of all (or atleast all popular) servers, because of the increased load.I suspect otherwise.

In summary, I believe the good of Netscape, far out-weighs any potential bad, and I heartily welcome an-other important WWW player.

Other BrowsersMany users either find Netscape more than theyneed, or are deterred by licensing concerns. Many


of them with Macintoshes or Windows machinesare turning to MacWeb and WinWeb from EINet(http://galaxy.einet.com/). These browsersare only in alpha release, and do have some problemswith HTML forms, but are otherwise very useful tools.There are several other commercial vendors of “en-hance” Mosaic, but I have not sampled these wares.

Mike MeyerCarnegie Mellon [email protected]

CONFERENCE NOTICES

Interface ’9527th Symposium on the InterfaceComputing Science and Statistics

June 21–24, 1995(note change of date)

Convention Center and Vista Hotel—Pittsburgh PA.

The Interface Conference is the premier annual confer-ence on the interface of computing and statistics. Itis sponsored by the non-profit Interface Foundation ofNorth America, and will be hosted in 1995 by CarnegieMellon University and the Pennsylvania State Univer-sity with Michael Meyer and James Rosenberger as jointprogram chairs. The 1995 conference will be in Pitts-burgh Pennsylvania, home of many excellent academicand industrial statistics and computer science researchprograms, and one of America’s most livable cities.

First Announcement and Call for PapersThe Symposium is being organized around the themeof “Statistics and Manufacturing,” with a sub-theme ofgreen manufacturing, the environment, and quantitativeenvironmental science.

The keynote speaker will be Raj Reddy, Dean of theSchool of Computer Science at Carnegie Mellon andthe most recent recipient of the A.M. Turing award.

Call for Papers

Sessions will include invited and contributed papers.Authors who wish to give a contributed paper shouldsubmit two copies of an abstract by February 28, 1995.The abstract should not exceed one-half page (text=6.5inches wide by 4 inches tall) with centered title, au-thor(s), and address. Invited and contributed papers areto be submitted for the proceedings by July 15, 1995.Proceedings submissions must be camera ready.

General InformationRegistration

The registration fee is $155 for members of the cooperat-ing societies, ASA, IMS, SIAM, ORSA, the BiometricsSociety (ENAR and WNAR) and for persons affiliatedwith Penn State University and Carnegie-Mellon Uni-versity. For others the fee is $175. The registrationfee covers the reception, coffee breaks, banquet, andproceedings.

Schedule

Conference registration begins on Wednesday after-noon (June 21). The first official conference event isa Wednesday evening mixer. Technical sessions willbe held on Thursday and Friday from 8:15 a.m.–5:15p.m. and from 8:15 a.m. to noon on Saturday. Breaksare scheduled between the sessions and also for lunch.All conference meetings will be held in the David L.Lawrence Convention Center, which is linked to thehotel by an enclosed walkway.

Entertainment

Baseball game. If you are a baseball fan, the Pitts-burgh Pirates will be hosting the San Francisco Giantson Wednesday, June 21st at 3:00 pm.

Thursday night banquet. This year’s banquet will takeplace on board one of the boats in the Gateway ClipperFleet. The three hour cruise will include sightseeingalong Pittsburgh’s three rivers, musical entertainment,and the Captain’s Buffet Dinner. The cost is includedas part of your registration fee.

Registration

For further registration and hotel information, contact:

Email: [email protected]: (412) 268-3108 Facsimile: (412) 268-7828Mail: Interface95, Department of StatisticsCarnegie Mellon University5000 Forbes Avenue, Pittsburgh, PA 15213, USA.

Further Information

For current information, use the World Wide Web toaccess the URL:http://www.stat.cmu.edu/interface95/

Mike MeyerCarnegie Mellon [email protected] L. RosenbergerPenn State [email protected]


SECTION OFFICERS

Statistical Graphics Section - 1994Roy E. Welsch, Chair

617-253-6601Massachusetts Institute of [email protected]

David W. Scott, Chair-Elect713-527-6037Rice [email protected]

Richard A. Becker, Past-Chair908-582-5512AT&T Bell [email protected]

William DuMouchel, Program [email protected]

Sallie Keller-McNulty, Program Chair-Elect913-532-6883Kansas State [email protected]

Michael M. Meyer, Newsletter Editor (93-96)412-268-3108Carnegie Mellon [email protected]

Linda A. Clark, Secretary/Treasurer (93-94)908-582-4807AT&T Bell [email protected]

Howard Wainer, Publications Officer (93-94)609-734-5623Educational Testing ServicePrinceton, NJ [email protected]

Jane F. Gentleman, Rep.(94-96) toCouncil of Sections613-951-8213Canadian Centre for Health InformationOttawa, ON K1A OT6, [email protected]

Sally C. Morton, Rep.(94-95) to Council of Sections310-393-0411The Rand CorporationSanta Monica, CA 90407-2138Sally [email protected]

Statistical Computing Section - 1994Trevor J. Hastie, Chair

908-582-5647AT&T Bell [email protected]

Mary Ellen Bock, Chair-Elect317-494-6053Purdue [email protected]

Sanford Weisberg, Past Chair612-625-8777University of [email protected]

Sallie Keller-McNulty, Program Chair913-532-6883Kansas State [email protected]

John A. Rice, Program Chair-Elect510-642-6930University of California at [email protected]

James L. Rosenberger, Newsletter Editor (93-96)814-865-1348The Pennsylvania State [email protected]

Deborah F. Swayne, Secretary-Treasurer908-829-4263Bell Communications ResearchMorristown, [email protected]

Karen Kafadar, Publications Liaison Officer301-496-8556National Cancer [email protected]

Ronald Thisted, Rep.(94-96) to Council of Sections312-702-8332/8333The University of [email protected]

Russell Lenth, Rep.(93-95) to Council of Sections319-335-0814University of [email protected]

Daryl Pregibon, Rep.(92-94) to Council of Sections908-582-3193AT&T Bell [email protected]


SECTION OFFICERS

Statistical Graphics Section - 1995David W. Scott, Chair

713-527-6037Rice [email protected]

William DuMouchel, Chair-Elect212-305-7736Columbia [email protected]

Roy E. Welsch, Past-Chair617-253-6601Massachusetts Institute of [email protected]

Sally C. Morton, Program Chair (appointed to replaceSallie Keller-McNulty)310-393-0411 ext 7360The Rand CorporationSally [email protected]

Stephen G. Eick, Program Chair-Elect708-713-5169AT&T Research [email protected]

Michael M. Meyer, Newsletter Editor (93-96)412-268-3108Carnegie Mellon [email protected]

Robert L. Newcomb, Secretary/Treasurer (95-96)714-824-5366University of California, [email protected]

Deborah J. Donnell, Publications Officer (94-96)206-283-8802 ext 258MathSoft, Seattle, WA

Colin R. Goodall, Rep.(95-97) to Council of Sections814-865-3993The Pennsylvania State [email protected]

Jane F. Gentleman, Rep.(94-96) toCouncil of Sections613-951-8213Canadian Centre for Health [email protected]

Sally C. Morton, Rep.(94-95) to Council of Sections310-393-0411 ext 7360The Rand CorporationSally [email protected]

Statistical Computing Section - 1995Mary Ellen Bock, Chair

317-494-6053Purdue [email protected]

Sallie Keller-McNulty, Chair-Elect913-532-6883Kansas State [email protected]

Trevor J. Hastie, Past ChairStanford [email protected]

John A. Rice, Program Chair510-642-6930University of California at [email protected]

Robert J. Tibshirani, Program Chair-Elect416-978-4642University of Toronto

James L. Rosenberger, Newsletter Editor (93-96)814-865-1348The Pennsylvania State [email protected]

Terry M. Therneau, Secretary-Treasurer507-284-1817Mayo [email protected]

Karen Kafadar, Publications Liaison Officer301-496-8556National Cancer [email protected]

MaryAnn H. Hill, Rep.(95-97) to Council of [email protected]

Michael M. Meyer, Rep.(95-96) Council of Sections412-268-3108Carnegie Mellon [email protected]

Ronald Thisted, Rep.(94-96) to Council of Sections312-702-8332/8333The University of [email protected]

Russell Lenth, Rep.(93-95) to Council of Sections319-335-0814University of [email protected]


INSIDE

A WORD FROM OUR CHAIRSStatistical Computing : : : : : : : : : : : : : : : 1Statistical Graphics : : : : : : : : : : : : : : : : : 1

FEATURE ARTICLETeaching Statistics and Computing via Multimedia

through the World Wide Web : : : : : : : : : 1EDITORIAL : : : : : : : : : : : : : : : : : : : : : 2FROM OUR CHAIRS (Cont.): : :

Statistical Graphics : : : : : : : : : : : : : : : : : 2Call for Ideas : : : : : : : : : : : : : : : : : : : : 3Statistical Computing : : : : : : : : : : : : : : : 4

STATISTICAL COMPUTINGStudent Paper Competition : : : : : : : : : : : : : 5

BUSINESS MEETING MINUTESStatistical Computing Section : : : : : : : : : : : 5Statistical Graphics Section : : : : : : : : : : : : 7

MIXER NOTESMixer at the Joint Meetings : : : : : : : : : : : : 8

CONTINUING EDUCATONStatistical Computing : : : : : : : : : : : : : : : 9

GRAPHICS SECTION RESOURCESGraphics Video library : : : : : : : : : : : : : : : 9

SECOND FEATUREThe Lisp-Stat Statistical Environment : : : : : : : 13

DEPARTMENTAL COMPUTINGEstablishing a Server : : : : : : : : : : : : : : : : 17

TOPICS IN SCIENTIFIC VISUALIZATIONA Colorful Variation On Box Plots : : : : : : : : : 19

NET SNOOPINGWWW viewers—Is there one for you? : : : : : : : 23

CONFERENCE NOTICESInterface ’95 : : : : : : : : : : : : : : : : : : : : 25

SECTION OFFICERS : : : : : : : : : : : : : : : : 26

The Statistical Computing & Statistical GraphicsNewsletter is a publication of the Statistical Comput-ing and Statistical Graphics Sections of the ASA. Allcommunications regarding this publication should beaddressed to:

James L. RosenbergerEditor, Statistical Computing SectionDepartment of StatisticsThe Pennsylvania State UniversityUniversity Park, PA 16802-2111(814) [email protected]

Michael M. MeyerEditor, Statistical Graphics SectionDepartment of StatisticsCarnegie Mellon UniversityPittsburgh, PA 15213-1380(412) [email protected]

All communications regarding membership in the ASAand the Statistical Computing or Statistical GraphicsSections, including change of address, should be sentto:

American Statistical Association1429 Duke StreetAlexandria, VA 22314-3402 USA(703) 684-1221 � FAX (703) [email protected]

Department of StatisticsUniversity Park, PA 16802-2111

Nonprofit OrganizationU. S. POSTAGE

P A I DPermit No. 1

University Park, PA 16802

Published by the Penn State Department of Statistics326 Classroom Building, University Park, PA 16802-2111

This publication is available in alternative media on request.Penn State is an affirmative action, equal opportunity university.U.Ed.SCI 95-87

Documents

StatisticalComputing TeachingStatisticsand ...stat-computing.org/newsletter/issues/scgn-05-3.pdffor obtaining assistance for learning and practicing sta-tistical methodology. ... [email protected]