THE ISBA BULLETIN
Vol. 10 No. 4 December 2003
The official bulletin of the International Society for Bayesian Analysis
APARTING MESSAGEFROMTHE each seen the stunning potential of Bayesian meth- PRESIDENT ods in our own work and experience. We are seeing the continued success and evolution of Bayesian by Ed George methods in theory and in applications across an ISBA President ever widening variety of fields. And yet surpris- [email protected] ingly, our membership has very diverse views on It has been a profound honor to serve as President what even constitutes Bayesian Analysis. Some of of ISBA for 2003. I would like to begin my part- us lean towards subjective purism, others towards ing Bulletin message by congratulating and wel- objective pragmatism. Some of us eschew any- coming our new leaders, President Elect Sylvia thing non-Bayesian, while others are eclectic, us- Richardson, Executive Secretary Deborah Ashby, ing Bayesian methods where and when it suits their and Board Members Brad Carlin, Merlise Clyde, needs. John Higdon and David Madigan. And a very But this diversity of views is our strength, es- special welcome to Heidi Sestrich who, I am de- pecially in our climate of camaraderie and mu- lighted to announce, has agreed to be the Admin- tual respect. The spirit of inclusiveness that per- istrative Director of ISBA. I would also like to bid vades ISBA provides an endless stream of energy a fond farewell to our departing Past President for debate and an environment where new ideas David Draper, Executive Secretary Cindy Chris- can easily rise and flourish. ISBA has become a tiansen and Board Members Nicky Best, Eduardo source of both intellectual stimulation and enor- Gutierrez-Pena, Tony O’Hagan and Raquel Prado. mous pleasure. The future of Bayesian Analysis Your dedication and hard work for ISBA have been looks brighter than ever, and we can all take great most gratefully appreciated by all. satisfaction in our contributions to nourishing this As ISBA’s 11th year comes to a close, ”We” can future through ISBA. We are the champions, and I all take great pride in how far ISBA has come. congratulate you all! We now regularly sponsor, cosponsor and endorse Bayesian meetings around the world. We oversee ' $ and award the four major Bayesian prizes, the De- Contents Groot, Lindley, Mitchell and Savage Awards. We produce the Bulletin that keeps Bayesians abreast ➤ ISBA 2004 of current developments. We are in the process of ☛ Page 2 creating Bayesian Analysis, a new flagship online ➤ Interview with C.Robert journal to attract and disseminate high impact, cut- ☛ Page 4 ting edge research developments. And I use ”We” because all of this has been a most remarkable joint ➤ Applications effort: The Executive Committee, The Board of Di- ☛ Page 8 rectors, The Program Council, The Prize Commit- tees, The Bulletin Editorial Staff, the newly assem- ➤ Annotated bibliography bled Bayesian Analysis Editorial Staff, the leaders ☛ Page 9 of the local ISBA Chapters in Brazil, Chile, India ➤ Student’s Corner and South Africa, and the myriad of members who ☛ Page 11 continue to participate in ISBA conference organi- zation and delivery. My deepest thanks and appre- ➤ Software review ciation to you all. I know of no other Society or ☛ Page 13 Association with such a large fraction of active and devoted contributors. ➤ News from the world ☛ Page 14 And why is this? I believe it is because we have & % ISBA Bulletin, December 2003 ISBA 2004
ISBA 2004 • Students members of ISBA: USD 75 VINADEL˜ MAR,CHILE • Students non-ISBAmembers: USD 85 MAY, 23-27, 2004 Registration fees for non-ISBA members include ANNOUNCEMENTS membership for 2004. After April 15th, 2004, regis- ➤ Scientific Programme tration fees will increase as follows: • ISBA members: USD 300 Lectures • non-ISBA members: USD 325 Opening Lecture Dani Gamerman, ”Dynamic spa- • Students members of ISBA: USD 125 tial models”. • Students non-ISBA members: USD 135 Closing Lecture Xiao-Li Meng, ”Sometimes it is possible to quantify ignorance: the case of Details on the payment of the registration fees Single Observation Unbiased Priors (SOUP)”. are available on the conference web site. Social Dinner Talk Jay Kadane, ”Bayesian statis- tics: the unfinished revolution”. ➤ ISBA Bulletin Tutorials: Introductory Level Hedibert Lopes ([email protected]), Edi- ”Bayesian Hierarchical Modeling”, David Draper. tor of the ISBA Bulletin, is organizing a meeting on ”Bayesian Model Choice”, Merlise Clyde. May, 24 at lunch time to discuss the Bulletin. He ”Bayesian Computation”, Peter Mueller. would like to receive comments and suggestions Tutorials: Advanced Level on how to improve the Bulletin.
”Decision Analysis in E-democracy”, Simon French. ➤ Deadlines ”Spatial Statistic”, Alan Gelfand and Sudipto Banerjee. ”Stochastic Optimization”, Christian Robert. • JANUARY 31, 2004. early submission of con- tributed posters Debates • FEBRUARY 15, 2004. notification of accep- Why don’t people use genuine prior information tance of early contributed posters in Bayesian applications? • APRIL 1, 2004. submission of contributed Speakers: Jim Berger, Phil Dawid, Pilar Igle- posters sias, Tony O’Hagan, Luis Raul Pericchi, Romano Scozzafava • APRIL 1, 2004. registration of all people pre- Chair: Jose-Miguel Bernardo senting or submitting papers for both oral On Bayesian Exploratory Data Analysis and poster sessions
Speakers: Susie Bayarri, Alicia Carriquiry, • APRIL 15, 2004. payment of registration fees Sid Chib, Hal Stern, Sergio Wechsler, at reduced rates Mike West Chair: Ed George ➤ Contacts ➤ Registration Fees For any information please contact the Chairs of the ISBA 2004 Committees: Registration fees for the ISBA 2004 meeting are listed below Scientific: Fabrizio Ruggeri [email protected] Finance: Alicia Carriquiry [email protected] • ISBA members: USD 200 Local: Pilar Iglesias [email protected] • non-ISBA members: USD 225
2 ISBA Bulletin, December 2003 ISBA 2004
ISBA 2004 at Vina˜ del Mar, on May 23-27, 2004. by Fabrizio Ruggeri For those of you who do not subscribe to [email protected] bayes-news, we have prepared the list of plenary speakers and the courses and the debates we have I am glad to report that our call for sessions and organized. student papers has generated many submissions, I remind you that the poster sessions are one of more than we expected. We got 34 session pro- the most relevant part of our conferences (for sure, posals and 28 student papers, besides some con- the most lively scientific activity, since I would not tributed posters and we are expecting many others! count the final act, the “usual” cabaret, as part of We are now reviewing the submitted abstracts and the scientific programme ...) Therefore, it is now the work is quite difficult since the papers we got time to submit papers for the poster sessions! We are, in general, very interesting. I wish to thank have decided to have two deadlines for submis- all those who contributed in organizing sessions sions: January 31st (for those willing to get a quick and those who accepted to be part of them. We answer, to be used to apply for funds) and April 1st were very pleased for the number and the quality (for all the others). of papers submitted by students. As I mentioned It is worth mentioning that all contributed pa- in my comments in the previous issue of the Bul- pers selected will automatically be considered for letin, we would like to have as many students as the Lindley Prize, awarded for innovative research possible. I have to thank ISBA officers (Ed George in Bayesian statistics, whereas selected papers pre- and Peter Mueller, first of all) and the Chairs of the sented at the conference (both oral and poster pre- Finance and Local Committees (Alicia Carriquiry sentations) will be published by the new electronic and Pilar Iglesias) since they agreed to financially ISBA journal, BAYESIAN ANALYSIS, following a support more students with respect to what we an- review process to be established by the Editorial nounced earlier. At this point, we would like to or- Board of the journal. ganize 6 student sessions and we can offer the stu- As you can see, we are preparing an exciting dents free lodging, the waiver of the registration fee programme with the help of the Bayesian fellows and the 2004 ISBA membership for free. Students worldwide and you can see how it takes shape by are applying from Europe, United States and South looking at the web site: http://isba.mat.puc.cl. America and one from Australia. You can find there details on accommodation, fees The proposed sessions cover a wide range of top- (I am sorry but we have to charge you something ics, with very scarce overlapping. The reading of ...), social events and, shortly, on accepted sessions the submitted abstracts is an interesting experience and student papers. to learn what is going on in the Bayesian world and If you have to decide which conference you do I am sure that the same feeling will be shared by not want to miss in 2004, my suggestion is ... the lucky ones who will attend ISBA 2004 in Chile
ISBA 2004 ways, very important. To this effect, we are prepar- by Pilar Iglesias ing a number of events that help making the most [email protected] out of the Conference and the natural beauty of Vina˜ del Mar. The local organizing committee is pleased to For further information about the ISBA 2004 invite the international scientific community to World Meeting, please point your browser to our participate in the ISBA 2004 World Meeting on web page http://isba.mat.puc.cl. Some infor- Bayesian Statistics, to be held inVinadel˜ Mar, Chile mation about the Valpara´ıso Region follows. in May 2004. This meeting is an opportunity to Valpara´ıso Region congregate Bayesian statisticians from all over the The pacific Ocean’s blue silhouettes the charms world, and to disseminate the discipline through- of two cities, Vina˜ del Mar, Chile’s Tourist Capital out Latin America. The Scientific Program includes and Valpara´ıso, World Heritage Site, Cultural Cap- various types of sessions as well as debates on var- ital and Legislative Seat. ious topics of current interest as well as poster ses- The intense mixture of history and legends are sions. ISBA is also particularly interested in pro- embodied in Valpara´ıso’s hills, they beckon you to moting the participation of young researchers and climb its narrow streets and special funicular eleva- students. tors and admire the beauty of a port , World Her- The social component of the meeting is, as al- itage Site, both for its architecture and its people.
3 ISBA Bulletin, December 2003 ISBA/INTERVIEW
This is a reason for artists, writers and musicians which is ideal for ecotourism fans. The Andes has been inspiring for its attractive. Pablo Neruda Mountains lay out their eternal snow for enjoying winner of Nobel reward, built charismatic homes the white sports or the generosity of its hot springs. in Valpara´ıso and Isla Negra. Vina˜ del Mar, is know And the wines give off their intense aromas, of- how the city yard, offers countless attractions and fering an unforgettable stroll through their routes entertainment that stand out among the beautiful while tasting their fine export vintage. flower gardens, and historic castles and buildings We want you to actively feel, see, taste, listen and and show room, where you constantly you can ap- live each one of the sensations that our Valpara´ıso preciate different activities of culture. Following Region offers you. We want you to experience that, the coastal road, its white sandy beaches and va- from the ocean to the mountains. riety gastronomy offer the promise of memorable Please, receive from LOC, which I preside, the outings. warmest welcome to this land that has a very mov- Nearby, the Cordillera Zone, is made up of ing nature and patrimony. the Cordillera de la Costa (Coast Mountain) and Sincerely, Pilar Iglesias, Chair of the Local orga- Cordillera de Los Andes (Andes Mountains). Beauti- nizing Committee. ful and unique landscapes stand out in this zone,
CHRISTIAN P. ROBERT lation: I was just used to work with one or two older colleagues who were reluctant to even pub- by Brunero Liseo lish technical reports. For fear of the ideas being [email protected] stolen or replicated or God knows what..!!! (Things have changed, I must say, and collaborations and Christian P. Robert is Professor of Statistics at the team work are not frown upon so much anymore.) CE.RE.MA.DE, Universite´ Paris IX. He is one of Another striking thing was the feeling of a com- the most prominent and active members of the munity, the sharing of many moments with the Bayesian community. He is the author of two rest of the faculty members and of the students. breakthrough books and many influential papers. There were lots of parties and dinners which were I sent questions to Chris by e-mail and here are his sort of blurring the distinction between work and responses. the rest, while, in France, the distinction between academic life and private life are fairly strict: no 1. Chris, you graduated in France and department picnic or Christmas party, no after- then moved to USA for Ph.D. studies. What seminar dinner either... During my Ph.D., I went were the most striking differences in the once for dinner at my advisor’s house and this academic life that you noticed between was quite exceptional as it never happened to my France and Cornell? friends. The part taken by Ph.D. students in the life of the department also was quite a new thing for me, because while in France, we were working Well, I actually did not move to the USA for Ph.D. in a rather strong isolation, sharing non-academic studies, but only after my Ph.D. I went first to Pur- moments with other Ph.D. students but very lit- due University in 1987, before spending one year tle in terms of research. Maybe again because of in Cornell and then moving back to France. I still this paranoia about stolen results... Anyway, these remember these two years very vividly because two years in the USA were essential for my future I then had the feeling of being part of something career and research: I learned how to share and very exciting. The most amazing thing, when com- to listen, to borrow and to communicate, and also ing from France (and especially the provincial uni- how to write papers, to work on several projects at versity I graduated from), was the ever-going de- a time, besides more basic skills like LAT X and C bate and movement and discussion about things: E programming. But, overall, I must say I keep from people were just excited doing research and ex- these two years a feeling of permanent fun and ex- changing ideas and starting collaborations! The citement. whole idea was simply too exhilarating!!! While, in France, there was much more secrecy and iso-
4 ISBA Bulletin, December 2003 INTERVIEW
2. France is certainly a country with a Das Gupta, George Strawderman, and great oth- strong mathematical tradition and culture: ers people I was to meet later, on the links between nonetheless, Bayesian ideas were not so James-Stein and Bayes, and found something much popular (although the recent studies of more to my taste. This is when I decided to spend Broemeling and Broemeling (Biometrika, some time in the USA to see the other side of it. 2003) and Broemeling (ISBA Bulletin, 2002) So, very logically, Econometrics drove me to work about E. Lhoste suggest differently). on James-Stein and James-Stein drove me to the How happened that you made the Bayesian Bayesian choice! choice ? 3. Do you still consider the decisionist It is true that Bayesian ideas were rather cen- step necessary to be a real Bayesian? I sored at the time I started my Ph.D.: Bayesian am experiencing the difficulty of teaching Statistics were not taught at all, there were no Bayes statistics having a small amount Bayesian talks during seminars or local conferences of time for introducing Bayesian ideas. and the very few people working on Bayesian Usually one cannot cover all the story. Statistics were not to be found in the Math and What is your opinion on that? Do you Stat departments but in Medical schools, Econo- consider possible to teach Bayes as if metric departments and Biometry research cen- it was just another brick to put on the ters. Although I was inclined towards Pure Math- likelihood wall ? ematics, the results of the entrance exams in the Grandes Ecoles drove me to a Statistics and Eco- I face the same difficulty when teaching, because nomics school, rather than the more prestigious the students are less and less exposed to the bases Ecole Normale Sup´erieure that I missed by a few of Decision Theory. At some point every Statis- points. (I had not applied to Ecole Polytechnique tics student would have read both of Lehmann’s because it was a military school!) So I spent the books, or some equivalent in French. With such a three years in this school (ENSAE) learning about basis and a very good understanding of losses, ad- Statistics, Econometrics and Economics, while pur- missibility and such, it was easy to step on Wald’s suing my Pure Mathematics dream by getting a theorem and the optimality of Bayes procedures. degree by correspondence at Paris 6 University. At Nowadays, if you want to use this rational presen- the end of these three years, I realized that Statis- tation, you first need to spend half of the course tics were much more interesting (for me) than Pure on Decision Theory, because the students are un- Math and looked around for both a Ph.D. advisor aware of this approach to Statistics. Given that I and a Ph.D. topic. Although I was quite tempted have the chance to teach the same students four to work with Dominique Picard on Differential Ge- years in a row, starting with their first Statistics ometry and Statistics, I eventually chose the topic class, I am actually privileged and can start pre- of James-Stein estimators, presumably because the senting the decisional bases of inference from the topic was more clearly defined and it had a link start. But in other programs, they already have had with the Econometrics classes I had had at ENSAE three years of Statistics and no Decision Theory and also because the advisor, Jean–Pierre Raoult, (and no Bayes either!) so I usually teach Bayesian seemed much more committed to take me as a Statistics through case studies with either known Ph.D. student than the other professors I had con- or unknown models, to show the students how tacted. I had heard very little of Bayesian Statistics much wider is the Bayesian approach. But I am not at the time, just 10 or 15 minutes in a Mathematical sure the students I am dealing with fully appreci- Statistics class loaded with disparaging comments, ate the distinction between likelihood and moment but I felt the idea of putting distributions on the (and Bayesian) estimators! parameters quite intriguing and provocative. Dur- ing my first Ph.D. year, Jean–Pierre Raoult gave 4. You have been (and you still are) me a draft of the book by Mouchart, Florens and one of the first Bayesians to rightly Rolin, Elements of Bayesian Statistics, stating that this consider the enormous impact of MCMC methods was a much better way of presenting the Bayesian in the development (even methodological) paradigm: I was then quite depressed because I of statistics. Can you describe us your thought it was much too measure-theoretic! But personal viewpoint on what is going to then I read all the papers by Jim Berger, Mary- be the future road of simulation based Ellen Bock, Larry Brown, George Casella, Anirban inference?
5 ISBA Bulletin, December 2003 INTERVIEW
Actually, I was privileged to attend one of theme from another domain. As they grow more Adrian Smith’s first lectures on the Gibbs sam- numerous and more expert, they will have a nat- pler, in June 1989 in Sherbrooke (Quebec,´ Canada), ural tendency to develop their own approach and and this was quite a shock. I can still visualize his sever the connection with statisticians. With the talk and the way he told us that with a few lines obvious danger of producing bogus Statistics... of code you could program the Gibbs sampler! To realize that “all” you had to do was to produce a 6. What kind of applications do you Markov chain to approximate any integral related like most? Were you ever involved in a to the stationary distribution...wow! Now, MCMC really exciting one? is part of the Bayesian toolbox and is taught even in elementary Bayesian classes. There is certainly less fundamental research going on in this area, even Applications usually come to me, rather than the though some recent developments about adaptive reverse, with a frequency of about once a year. I MCMC algorithms seem quite exciting (to me!). started getting interested in mixtures because my As the Bayesian community is growing, we see wife was working on her Master thesis in Ottawa other approaches to this computational problem, and was analyzing radiographs with bimodal fea- with various degrees of approximation. Take for tures. Since then I collaborated to applications in instance the machine learning community and the astronomy, drug testing, genetics, econometrics variational approximations: they are quite interest- and engineering. I really like the astronomy ap- ing alternatives, even though they seem to depend plications because astronomers, being quite close highly on the problem at hand. I am also quite to mathematicians, lay things very neatly and it is eager to see how the particle system (or sequential thus much easier to process their problems. Right Monte Carlo) community is going to evolve, be- now, I am involved in the analysis of a survey of cause they bring somehow more efficient solutions treatments of kids with a rare disease called Evans to the problems tackled by MCMC algorithms. syndrome and this is both exciting and challenging Further in the future, I do not want to make pre- at once. dictions: certainly, improved computer technology will allow for more exhaustive methodologies to be 7. I have always seen you very active used, but, at the same time, the complexity of the during the conferences, doing trekking, problems considered will increase in parallel. And skiing and many other things. Tell us I am not sure computing is the number one prob- about how do you like to spend your time, lem when handling fuzzier problems like those in when you are not dealing with statistics very large dimensions: modeling comes first. or ....kids!
5. You are a mixture expert. Please My kids indeed take a lot of my free time, be- use your favorite MCMC methods to estimate sides Statistics, but they also bring a lot to my life! the number of components (and the weights As you noticed, I do try to take advantage of ev- of them) which are going to compose the ery opportunity to go mountaineering in France profile of the next generation’s statistician.or abroad, and I must say this creates sort of a bias in my conference planning. Skiing is very re- cent and I only started it because I happened to This is a hard one, too. I am actually very afraid organize a Winter Research Kitchen in a ski resort. that my MCMC sampler could diverge to infinity Besides these too rare trips to the mountains, I do for this estimation or, worse, to 0. There somehow a lot of running in the Summer, training for a ritual is a danger of assimilation of Statistics by the users, half-marathon in the Fall. (Responsability for get- with a tiny weeny core of statisticians resisting to ting me started into that is shared between George the lure of applications and joining probabilists. Casella and Peter Green.) And I also like reading It is somehow the (negative) consequences of the a lot, my eclectic taste being somehow reflected in success of Statistics in that, more and more, we the quotes used in my books. find people using (well or not) Statistics with main
6 ISBA Bulletin, December 2003 INTERVIEW
8. You organized the last "‘Objective pedagogical. Bayes" Conference. Do you really believe that this sort of foundational compromise 10. What is and what should be, in your can be successful for the future of our opinion, the role of European statistician discipline? within ISBA? Are you in favor of a kind of European chapter of the society? This is not clear to me, but I found that there was renewed interest for Bayesian Statistics in This is quite an astonishing question, as I was some communities that I first considered as “too” thinking about a European structure subsuming mathematical. The conference in Aussois was, be- our national societes this very morning, while run- sides a reunion of hard-core objective Bayesians!, ning! I am uncertain whether or not chapters and an attempt to bring these communities closer to such societal structures can help to improve the mainstream Bayesian Statistics. I am not sure the visibility of Bayesians and their networking, but meeting was a success in this respect! At another there is a lot to be said about EU programs like the level, there always is a need for deeper founda- TMR (Travel and Mobility of Researchers) program tions and, if we cannot keep the link with Decision that contributed a lot to the advances of MCMC Theory, I feel that asymptotic, minimax and other and to the creation of a closely linked community frequentist-Bayes arguments may help. So for me around MCMC. In some European countries, there this is not a “compromise”. Frequentist validation is not enough of a tradition of moving around and may be a necessary counter-power to the uncon- we should take advantage of every opportunity trolled expansion in the use of data-dependent and to get young Ph.D.’s for a year or two in another other ad hoc priors. center abroad. Student traveling programs should also be used to start a sort of European (Bayesian) 9. Is there any other chapter you would Ph.D. program, where students would move from like to add to your Bayesian Choice ? one university to another to get his/her degree and a much better training in Bayesian Statistics. Noel Cressie once told me I should add one Last week I attended a European Young Statisti- chapter on spatial Statistics... In fact, the book cian meeting in Switzerland (as the old guy!) and would need a sequel containing applications to I found quite exceptional the reunion of all these many fields (including spatial Statistics, of course). young statisticians with so diverse experiences and Now, I would have much more trouble writing visions of what Statistics is and should be. If we this sequel than the Bayesian Choice and so I do could share these visions on a broader scale this not think I ever will. A particular application that would help the field dramatically. Of course, I could (maybe!) justify a new chapter would be a did not really answer the question, since I rather thorough processing of generalized linear models, addressed the problem of European statisticians that would come as an application of both MCMC as a whole. But, since the 21st Century will be and hierarchical models chapters. I would also Bayesian... like to find a radically new way of presenting and handling the testing chapter, as I think the way it Thanks to Chris for his stimulating and is presented in the current version is not the most informative answers.
SUGGESTIONS PLEASE, FEELCOMPLETELYFREETOSENDUSSUGGESTIONSTHATMIGHT IMPROVETHEQUALITYOFTHEBULLETIN [email protected]
7 ISBA Bulletin, December 2003 APPLICATIONS
WAVELET-BASED stage of carcinogenesis, i.e. the first few hours after NONPARAMETRIC MODELINGOF the carcinogen exposure. In the experiment 30 rats HIERARCHICAL FUNCTIONSIN were fed one of the 2 diets for 14 days, exposed to COLON CARCINOGENESIS a carcinogen, then sacrificed at one of 5 times af- ter exposure (0, 3, 6, 9, or 12 hours). Rat’s colons by Marina Vannucci were removed and dissected, and measurements [email protected] of various biomarkers, including MGMT, were ob- tained. Each biomarker was measured on a set of The September issue of the Journal of the American 25 crypts in the distal and proximal regions of each Statistical Association features the Bayesian work rat’s colon. Crypts are fingerlike structures that ex- of Morris, Vannucci, Brown and Carroll, Wavelet- tend into the colon wall. The procedure yielded Based Nonparametric Modeling of Hierarchical Func- observed curves for each crypt consisting of the tions in Colon Carcinogenesis, as the Applications biomarker quantification as a function of relative and Case Studies Invited Paper for 2003. This cell position within the crypt, the position being re- work has also been selected as the winner of the lated to cell age and stage in the cell cycle. Due to 2003 Mitchell Prize, awarded annually in recogni- the image processing used to quantify the measure- tion of an outstanding paper that describes how a ments, these functions may be very irregular, with Bayesian analysis has solved an important applied spikes presumably corresponding to regions of the problem. crypt with high biomarker levels. Most of the published work in wavelet nonpara- The primary goal of the study was to determine metric regression is limited to the single curve set- whether diet has an effect on MGMT levels, and ting. In their work, Morris, Vannucci, Brown and whether this effect depends on time and/or rela- Carroll extend wavelet regression to the frame- tive depth within the crypt. Another goal was to work of nested functional data. Their Bayesian assess the relative variability between crypts and method leads to adaptively regularized estimates between rats. and posterior credible intervals for the mean func- In their work, Morris, Vannucci, Brown and Car- tion and random effects functions, as well as the roll apply their hierarchical model to the func- variance components of the model. The approach tional data from this colon carcinogenesis study first applies a discrete wavelet transform (DWT) to and fit the model using an MCMC procedure to ob- each observed curve to obtain the corresponding tain posterior samples from the functions at each wavelet coefficients. This step results in the projec- hierarchical level and the variance components. tion of the original curves into a transformed do- The function estimates at all levels are adaptively main, where modeling can be done in a more par- regularized using a multiple shrinkage prior im- simonious way. A Bayesian model is then fit to posed at the top level of the hierarchy. The treat- each wavelet coefficient across curves. The inverse ment level functions are directly regularized by DWT is finally applied to transform the obtained this shrinkage, while the functions at the lower estimates back to the data domain. The method is levels of the hierarchy are subject to some regu- particularly suitable to modeling hierarchical spa- larization induced by the higher levels, as mod- tially heterogeneous functional data that are char- ulated by the variance components. The authors acterized by local features, such as peaks. provide guidelines for selecting these regulariza- The work was motivated by a case study inves- tion parameters, together with empirical Bayes es- tigating the effect of diet on O6 -methylguanine- timates, introduced during the rejoinder to the dis- DNA-methyltransferase (MGMT), an important cussion. biomarker in early colon carcinogenesis that mea- Results from the analyses of the case study reveal sures the amount of repair enzyme. Dietary factors, that there is more MGMT expressed at the lumenal such as the amount and type of fat and fiber, have surface of the crypt, and suggest a diet difference been shown to be related to the risk of colon can- in the MGMT expression at this location 12 hours cer. Data from animal experiments are often col- after exposure to the carcinogen. Also, the mul- lected to investigate the biological mechanisms be- tiresolution wavelet analysis highlights features hind carcinogenesis. present in the crypt-level profiles that may corre- In the specific case study analyzed by Morris, spond to individual cells, suggesting the hypoth- Vannucci, Brown and Carroll two types of dietary esis that MGMT operates on a largely cell-by-cell fat (fish oil or corn oil) are investigated as po- basis. tentially important factors that affect the initiation
8 ISBA Bulletin, December 2003 APPLICATIONS/ANNOTATED BIBLIOGRAPHY
Reference: parametric Modeling of Hierarchical Functions in Colon Carcinogenesis (with discussion). Journal of MORRIS,J.S.,VANNUCCI,M.,BROWN, P.J. the American Statistical Association, 98, 573–597. and CARROLL,R.J. (2003). Wavelet-Based Non-
COVARIANCE ESTIMATION • Bernardo, J. M. & Smith, A. F. M. (1994). Bayesian Theory. New York: John Wiley & by Merrill W. Liechty Sons, Inc. [email protected] The most commonly used prior model is the conjugate inverse-Wishart. It allows closed Any collection of multivariate data has a covari- form posterior inference, and efficient im- ance matrix to be estimated. This is not always plementation of Gibbs sampling schemes for straight forward due to several complex charac- more complex models with additional pa- teristics of the covariance matrix that have tradi- rameters beyond the unknown covariance tionally been inhibitive. With increased comput- matrix. However, this prior model has the ing power available today however, many of these drawback that there is a single degree of free- concerns are no longer restrictive. At the heart of dom parameter, which is the only “tuning pa- covariance estimation are the issues of interpreta- rameter” available to express uncertainty. tion and prior elicitation, the (potentially) restric- tively large size of the matrix, and finally the space • Bollerslev, T. (1986). Generalized autoregres- of the positive definite matrix. It is important to sive conditional heteroscedasticity. Journal of have models that take into account these issues and Econometrics. 31, 307–27. allow the researcher to accurately model the cor- The author extends the ARCH model to the relation structure. Departures from standard MLE GARCH model by including lagged values and inverse-Wishart prior models have recently be- of the variance itself in the variance equa- come popular. tion thereby providing a more flexible model form. • Aguilar, O. & West, M. (2000). Bayesian dy- namic factor models and portfolio allocation. • Bollerslev, T., Engle, R. F. & Wooldridge, M. Journal of Business and Economic Statistics. 18, (1988). A capital asset pricing model with 338–57. time varying covariances. Political Economy. 96, 116–31. The authors propose Bayesian factor models which are natural candidates for correlation A multivariate version of the GARCH model estimation. Factor models effectively reduce is presented by the authors. The high dimen- the dimensionality of the covariance matrix. sional nature of the covariance matrix com- plicates modeling.
• Barnard, J., McCulloch, R. & Meng, X. (2000). • Bollerslev, T. (1990). Modelling the coherence Modeling covariance matrices in terms of in short-run nominal exchange rates: A mul- standard deviations and correlations, with tivariate generalized ARCH approach. Re- applications to shrinkage. Statistica Sinica. 10, view of Economics and Statistics. 72, 498–505. 1281–1311. The author adapts the GARCH framework The authors use a separation strategy for for multivariate models by using a time vary- modeling Σ = SRS by assuming indepen- ing conditional covariance matrix with time dent priors for the standard deviations S and invariant conditional correlations. the correlation matrix R. They propose two alternative prior models for R. One is the • Chib, S. & Greenberg, E.(1998). Analysis of marginally uniform prior, where marginally multivariate probit models. Biometrika. 85, each correlation can be uniform over [−1, 1]. 347–61. The other is the jointly uniform prior, where The authors discuss the constraint to posi- the matrix R is assumed to be a priori uni- tive definiteness and the typically high di- formly distributed over all possible correla- mensional nature of the parameter vector for tion matrices. the covariance matrix.
9 ISBA Bulletin, December 2003 ANNOTATED BIBLIOGRAPHY
• Daniels, M. J.(1999). A prior for the variance and allows for hierarchical shrinkage to be in hierarchical models. Canadian Journal of done with the eigenvalues. Statistics. 27, 567–78. • Liechty, J. C., Liechty, M. W., & Muller,¨ The author proposes a uniform shrinkage P. (2004). Bayesian Correlation Estimation. prior for the covariance matrix. Biometrika. 91. to appear. • Daniels, M. J. & Kass, R. E. (1999). Nonconju- The authors put mixture priors on the corre- gate Bayesian estimation of covariance matri- lations allowing the correlations or the vari- ces and its use in hierarchical models. JASA. ables to group in natural ways. They deal 94, 1254–63. with problems that arise from the truncation The authors discuss three alternative hierar- of priors caused by the positive definite con- chical priors. The first is a hierarchical exten- straint. sion of the inverse-Wishart prior. In the sec- ond they assume a normal prior for a trans- • Lin, S. P. & Perlman, M. D. (1985). An im- formation of the correlation coefficients. The proved procedure for the estimation of a cor- third model uses an eigenvalue/eigenvector relation matrix. In Statistical Theory and Data parameterization, with the orthogonal eigen- Analysis, Ed. K. Matusita, pp. 369–79. North vector matrix parameterized in terms of the Holland: Elsvier. Givens angles. The original frequentist work using a version of the James-Stein estimator to model the off • Daniels, M. J. & Kass, R. E. (2001). Shrink- diagonal elements of the correlation matrix. age estimators for covariance matrices. Bio- metrics. 57, 1174–84. • West, M. (2003). Bayesian factor regression The authors use a Fisher z transformation of models in the “large p, small n” paradigm. In the correlations. Bayesian Statistics 7. Eds. J.O. Bernardo et al., pp. 723-32, Oxford. • Engle, R. F. (1982). Autoregressive condi- tional heteroscedasticity with estimates of the The author proposes Bayesian factor models variance of United Kingdom inflation. Econo- which are natural candidates for correlation metrica. 50, 987–1008. estimation. The author introduces ARCH (autoregres- • Wong, F., Carter, C. K. & Kohn, R. (2003). Effi- sive conditional heteroskedasticity) models, cient estimation of covariance selection mod- which have achieved a considerable follow- els. Biometrika. 90. pp. 809-830. ing in the econometrics and finance literature. The authors propose a prior probability • Karolyi, G. A. (1993). A Bayesian approach model on the precision matrix (P = Σ−1) that to modeling stock return volatility for option focuses on the sparseness of the precision ma- valuation. Journal of Financial and Quantitative trix, and is geared towards graphical models Analysis. 28, 579–94. and partial correlations. They interpret zeros The author uses Bayesian methods to esti- in the partial correlation matrix in the graphi- mate the variance of individual stock returns cal models context as meaning conditional in- based on stocks grouped a priori according to dependence. size, financial leverage, and trading volume. • Yang, R. & Berger, J. O. (1994). Estimation of • Leonard, T. & Hsu, J. S. (1992). Bayesian infer- a covariance matrix using the reference prior. ence for a covariance matrix. Annals of Statis- Annals of Statistics. 22, 1195–211. tics. 20, 1669–96. The authors propose a reference prior for the The author introduces the log matrix prior covariance matrix that is inversely propor- by using a logarithmic transformation of the tional to the product of differences in ordered eigenvalue/eigenvector decomposition of Σ eigenvalues of the covariance matrix Σ.
10 ISBA Bulletin, December 2003 STUDENT’S CORNER
SURVEYON BAYESIAN PAPERS Berger, J. O. (2000) ”Bayesian analysis: A look at today and thoughts of tomorrow”. Journal of by Lilla Di Scala and Luca La Rocca the American Statistical Association, 95, pp. [email protected] [email protected] 1269-1276.
In this issue, we report on the results of the Berger, J. O. (2003) ”Could Fisher, Jeffreys and ”naive” survey which we had launched in a past Neyman have agreed upon testing?”. Statis- issue of the Bulletin. The question we put forward tical Science, 18, pp. 1-32. was ”Which groundbreaking/essential papers do Berger, J. O. and Bernardo, J. M. (1992) ”On the de- you believe that a graduate student with a serious velopment of reference priors”. In Bayesian interest in Bayesian statistics should not miss read- Statistics 4 (J. M. Bernardo et al., eds.), pp. 61- ing?” 77, Oxford University Press. In particular, let us stress that the results rep- resent the opinions of both those who replied to Berger, J. O. and Delampady, M. (1987) ”Testing our call and those who were contacted directly via precise hypotheses”. Statistical Science, 2, pp. email, as part of a ”hand-made” list of Bayesians. 317-352. Unfortunately, we did not receive enough replies Bernardo J. M (1979) ”Reference posterior distri- to set up a ”decent” statistical analysis of the re- butions for Bayesian inference”. Journal of sults. We would be grateful if someone would like the Royal Statistical Society, Series B, 41, pp. to point out to us the reasons why our question 113-147. did not provoke a wider response. However, we believe that truly interesting suggestions came up Bernardo, J. M. (1997) ”Noninformative priors do and we take this opportunity to thank all of those not exist: a discussion”. Journal of Statistical who participated and gave up some of their time Planning and Inference, 65 , pp. 159-189. to formulate an annotated list of their all-time fa- vorites. Comments, additions, and constructive Bernardo, J. M. and Rueda, R. (2002) ”Bayesian complaints are more than welcome, as the list here- hypothesis testing: A reference approach”. under is by no means complete. Only papers have International Statistical Review, 70 , pp. 351- been taken into account: no good books and no 372. links to interesting Web pages have been consid- ered, even if some readers suggested them directly. Bernardo, J. M. (2003) ”Bayesian Statistics”. In fact, we deemed better to stick to our original In Encyclopaedia of Life Support Systems idea, which was for you all to help PhD students (EOLSS), Paris: UNESCO (in press). find their course in the vast ocean of research pa- Besag, J. (1974) ”Spatial interaction and the statis- pers. tical analysis of lattice systems”. Journal of The below references are to the first appearance the Royal Statistical Society, Series B, 36, pp. of a paper and they are kept short due to space re- 192-236. strictions. Some people mentioned their own work (”the sum of many subjective choices will make Birnbaum, A. (1962) ”On the Foundations of Sta- a sensible representative list”), some other people tistical Inference”. Journal of the American preferred old and well-established papers (”only Statistical Association, 57, pp. 269-306. historical perspective can tell you if a paper is in- deed essential”). We report ALL mentioned pa- Dawid, A. P. (1979) ”Conditional independence in pers, old and new, by author, in alphabetical order. statistical theory”. Journal of the Royal Statis- Top scoring papers are Lindley and Smith (1972), tical Society, Series B, 41, pp. 1-31. Gelfand and Smith (1990) and de Finetti (1937). Let us also mention repeated votes for Berger (2000) de Finetti, B. (1931) ”Sul significato soggettivo and Tierney (1994). All other papers were sug- della Probabilita”.` Fundamenta Matemati- gested only once, which makes for a longer list of cae, 17, pp. 298-329 (in Italian). interesting papers: enjoy! de Finetti, B. (1937) ”La prevision:´ ses lois Berger, J. (1984) ”The robust Bayesian view- logiques, ses sources subjectives”. Annales point”. In Robustness in Bayesian Statistics de l’Institut Henri Poincare,´ 7, pp. 1-68 (in (J. Kadane ed.), pp. 63-124, North-Holland. French).
11 ISBA Bulletin, December 2003 STUDENT’S CORNER
Edwards, W., Lindman, H., and Savage, L. J. Kennedy, M. C. and O’Hagan, A. (2001) ”Bayesian (1963) ”Bayesian statistical inference for psy- calibration of computer models”. Journal of chological research”. Psychological Review, the Royal Statistical Society, Series B, 63, pp. 70, pp. 193-242. 425-464.
Ferguson, T. S. (1973) ”A Bayesian analysis of Lauritzen, S. L. and Spiegelhalter, D. J. (1988) ”Lo- some nonparametric problems”. Annals of cal computations with probabilities on graph- Statistics, 1, pp. 209-230. ical structures and their application to ex- pert systems” (C/R: p194-224). Journal of the Gelfand, A. E. and Smith, A. F. M. (1990) Royal Statistical Society, Series B, 50, 157-194. ”Sampling based approaches to calculating marginal densities”. Journal of the American Lindley, D. V. (2000) ”The philosophy of statis- Statistical Association, 85, pp. 398-409. tics”. The Statistician, 49, pp. 293-337.
Gelman, A. and Rubin, D. (1992) ”Inference Lindley, D. V. and Novick, M. R. (1981) ”The from iterative simulation using multiple se- role of exchangeability in statistical infer- quences.” Statistical Science, 7, pp. 457-472. ence” The Annals of Statistics, 9, pp. 45-58.
Goldstein, M. (1999) ”Bayes linear analysis”. In Lindley, D. V. and Phillips, L. D. (1976) ”Inference Encyclopaedia of Statistical Sciences (Kotz, S. for a Bernoulli process (A Bayesian view)”. et al.), update volume 3, pp. 29-34, Wiley. The American Statistician, 30, pp. 112-119.
Green, P. (1995) ”Reversible jump Markov chain Lindley, D. V. and Smith, A. F. M. (1972) ”Bayes es- Monte Carlo computation and Bayesian timates for the linear model”. Journal of the model determination”. Biometrika, 82, pp. Royal Statistical Society, Series B, 34 , pp. 1- 711-732. 41.
Harrison, P. J. and Stevens, C. F. (1976) ”Bayesian Lindley, D. V., Tversky, A. and Brown, R.V. (1979) forecasting” (C/R: pp. 228-247). Journal of ”On the reconciliation of probability assess- the Royal Statistical Society, Series B, 38 , pp. ments”. Journal of the Royal Statistical Soci- 205-228. ety, Series A, 142, pp. 146-180. Richardson, S. and Green, P. (1997) ”On Bayesian Imbens, G. and Rubin, D. (1997) ”Bayesian Infer- analysis of mixtures with an unknown num- ence for causal effects in randomized exper- ber of components”. Journal of the Royal Sta- iments with noncompliance”. The Annals of tistical Society, Series B, 59 , pp. 731-792. Statistics, 25, pp. 305-327. Rubin, D. B. (1976) ”Inference and missing data”. Kadane, J. B. (1980) ”Predictive and structural Biometrika, 63, pp. 581-592. methods for eliciting prior distributions”. In Bayesian Analysis in Econometrics and Rubin, D. B. (1978) ”Bayesian inference for causal Statistics: Essays in Honor of Harold Jeffreys effects: The role of randomization”. The An- (Zellner, A. ed.), pp. 89-109, North-Holland. nals of Statistics, 7, pp. 34-58.
Kadane, J. B. and Wolfson, L. J. (1998) ”Experi- Rubin, D. B. (1984) ”Bayesianly justifiable and ences in elicitation”. The Statistician, 47, pp. relevant frequency calculations for the ap- 1-20. plied statistician”. Annals of Statistics, 12, pp. 1151-1172. Kass, R. E. and Raftery, A. E. (1995) ”Bayes fac- tors”. Journal of the American Statistical As- Savage, L. J. (1971) ”The elicitation of personal sociation, 90, pp. 773-795. probabilities and expectations”. Journal of the American Statistical Association, 66, pp. Kass, R. E. and Wasserman, L. (1996) ”The selec- 783-801. tion of prior distributions by formal rules”. Journal of the American Statistical Associa- Shafer, G. (1986) ”Savage revisited”. Statistical tion, 91, pp. 1343-1370. Science, 1, pp. 463-501.
12 ISBA Bulletin, December 2003 STUDENT’S CORNER/SOFTWARE REVIEW
Smith, A. F. M. and Gelfand, A. E. (1992) and Cowell, R. G. (1993) ”Bayesian analysis ”Bayesian statistics without tears: A of expert systems”. Statistical Science, 8, pp. sampling-resampling perspective” (Corr: 219-247. 93V47 p. 158). The American Statistician, 46 , pp. 84-88. Tierney, L. (1994) ”Markov chains for exploring posterior distributions”. The Annals of Statis- Spiegelhalter, D. J., Freedman, L. S. and Parmar, tics, 22, pp. 1701-1762. M. K. B. (1994) ”Bayesian approaches to ran- domized trials”. Journal of the Royal Statisti- West, M., Harrison, J., and Migon, H. (1985) cal Society, Series A, 157, pp. 357-416. ” Dynamic Generalized Linear Model and Bayesian Forecasting”. Journal of the Ameri- Spiegelhalter, D. J., Dawid, A. P., Lauritzen, S. L., can Statistical Association, 80, 73-97.
BAYESIAN ANALYSIS OF Whereas, to justify the model output to manage- SOFTWARE COST ESTIMATING ment (for example, to justify the cost estimate for MODEL:COCOMOII project management or for bidding on a competi- tive proposal), Model D is preferred. Lets say, you by Sunita Chulani have a new project to estimate. Model D estimates [email protected] it will take 200 person-months (PM) to complete and Model E estimates an effort of 100 PM. Which Software cost estimation has become an increas- model’s output should you believe? ingly important field due to the pervasive role of Should you present the average estimate of 150 software in today’s world. In spite of the existence PM? Is the average justifiable? A preferred ap- of about a dozen software estimation models, the proach is one that balances the results of Model D field continues to remain not-too-well-understood, and Model E, such that it favors the experts for pa- causing growing concerns in the software mea- rameters where they are in strong agreement and surement/engineering community. This article de- favors the data fit for parameters where the data scribes the use of the Bayesian methodology on a is strong and the experts disagree i.e. a mathe- well-known non-proprietary software cost model: matical technique for creating a Model B (Balanced COCOMO II. Lets say you have two software cost or Bayesian). In the COCOMO II version of the estimation models: Bayesian approach, the Model E parameter values 1. A data-determined model, Model D that has and their variances are taken as the a priori knowl- all its parameters calibrated to the real soft- edge about the parameter values. The Model D pa- ware project data. Since software project data rameter values and their variances are then taken is inevitably collected with imprecise defini- as new information, which can be used to deter- tions of what is included in the product, in mine an a posteriori update of the parameter val- the process, and in the workforce, some of the ues. The Bayesian approach produces a weighted parameters calibrated to the noisy data are average of the Model D and E values, which gives counter intuitive i.e. do not agree with expert higher weights to parameter values with smaller judgment. variances. A tool based on Model B can be ob- tained from the Center for Software Engineering 2. An expert judgment based model, Model E. at University of Southern California’s website: It has the same parametric form and parame- http://sunset.usc.edu/research/COCOMOII. ters as Model D, but it is calibrated using ex- More details on the Bayesian calibrated CO- pert consensus. COMO II can be obtained in Boehm et al’s Mathematically, Model D will always have a better Prentice Hall book, ”Software Cost Estimation fit to project data than Model E. But, many experts with COCOMO II”, ISBN 0-13-026692-2 and in in the field of software estimation strongly believe Chulani’s Ph.D. dissertation, ”Bayesian Anal- that Model E captures software processes and phe- ysis of Software Cost and Quality Models”, nomena better than Model D, and is a better choice. http://sunset.usc.edu/publications.
13 ISBA Bulletin, December 2003 NEWS FROM THE WORLD
NEWSFROMTHEWORLD Khuri, Clyde Schoolfield, and Bhramar Mukherjee. http://www.stat.ufl.edu/symposium/2004/dmbio by Gabriel Huerta [email protected] SPRUCE Advanced Workshop on Spa- ✽ denotes an ISBA activity tial/Temporal Models and Methods March 24-27, 2004. Costa Do Estoril, Lisbon ➤ The workshop will be on the theme Environ- Events mental Problems an will be concerned, inter alia, with the following applications: traditional en- vironmental modeling, climatology and meteo- One day workshop on Bioinformatics and a rology, epidemiology and health, physical sys- Bayesian afternoon December 4-5, 2003, Guanaju- tems, biology and agriculture, and modeling prin- ato, Mexico ciples: heuristic/empirical ad hoc approaches, Invited Speakers: Keith Baggerly, (MD An- harmonic frequency-based models, kriging and derson Cancer Center, Houston) Kim-Anh Do, Kalman filter, space-time point process models (MD Anderson Cancer Center, Houston) Jun and hierarchical Bayesian techniques. Leading Liu, (Harvard) Peter M¨ller, (MD Anderson Can- international speakers are expected to include cer Center, Houston) Mike West, (Duke). Orga- Paul Blackwell, Patrick Brown, Noel Cressie, Pe- nizer: J. Andres´ Christen. More information at ter Diggle, Tilmann Gneiting, Gudmund Host, http://www.cimat.mx/talleres/Bioinformatics Moira Mugglestone, Doug Nychka, Sujit Sahu, and Jim Zidek. Contact Vic Barnett, School ✽ IMS-ASA’s SRMS Joint Mini Meeting on of Computing and Mathematics, Nottingham Current Trends in Survey Sampling and Official Trent University, [email protected] Web page: Statistics January 2-3, 2004,Calcutta, India http://spruce.deio.fc.ul.pt Co-sponsors: Department of Statistics, Univer- sity of Calcutta. Gallup Research Center, Univer- CART Data Mining 2004 March 22-24, 2004. sity of Nebraska-Licoln, U.S. Census Bureau. Main Conference San Francisco. May 11-12, 2004. Endorsed by: International Society for Bayesian Madrid conference Analysis (ISBA) First international CART conferences focusing Contact Partha Lahiri at [email protected] on the Data Mining technology of Leo Breiman if you are interested in presenting your paper (U.C. Berkeley), Jerome Friedman (Stanford), at the meeting. For local information, please Richard Olshen (Stanford) and Charles Stone (U.C. contact Tathagata Banerjee at [email protected]. Berkeley). The conference series will provide an http://www.jpsm.umd.edu/ims opportunity for data mining professionals to ex- University of Florida Sixth Annual Winter change ideas on the art and practice of the real Workshop Data Mining, Statistical Learning, and world analysis of complex data. Contributed pa- Bioinformatics January 8-10, 2004. University of pers covering any application of CART, MARS, Florida - Department of Statistics PRIM, and TreeNet are encouraged, including in- The Department of Statistics at the University novative and unusual applications. The confer- of Florida will host its Sixth Annual Winter Work- ences are intended to serve several functions. A shop in Gainesville, Florida during January 8-10, celebration of technology to honor the four authors 2004. The workshop will focus on recent develop- of CART, with the opportunity to meet with them ments in statistical methods for analyzing massive, in person. Each is planning to offer a keynote pa- unorthodox datasets and on applications of such per. A venue to exchange ideas and experiences techniques in many areas. A major purpose of the focused on the practice of data mining. A net- workshop is to foster interaction between senior working opportunity leading to the creation of and young researchers, to assess the current state local use groups and the establishment of a user of art in data mining and bioinformatics and to newsletter. A place to learn about extensions to promote discussion in related topics. In addition to CART related technology and anticipated future invited presentations, the symposium will include developments. An opportunity to obtain both a contributed poster session. The symposium is or- basic and advanced training offered by practical ganized by George Casella, Jim Booth, Jim Hobert, and theoretical experts. For more information visit Alan Agresti, Malay Ghosh, Mike Daniels, Andre http://www.cartdatamining.com
14 ISBA Bulletin, December 2003 NEWS FROM THE WORLD
International Conference on Analysis of Ge- nomic Data May 10-11, 2004. The Conference Center at Harvard Medical School The conference is jointly organized by the Boston ✽ The IV Workshop on BAYESIAN NON- Chapter of the American Statistical Association PARAMETRICS: Metodology, Theory and Appli- (BCASA) and the Channing Laboratory in the De- cations June 13-16, 2004 at the Conference Center of partment of Medicine of the Brigham and Women’s Universit`a ”La Sapienza” via Salaria 113 - Roma - Hospital and Harvard Medical School. ITALY. AIM and SCOPE The Workshop will cover the following broad The conference will cover current research on areas: Foundations and Prior Construction, Mod- statistical models and bioinformatic methods for eling, Asymptotics, Computation and Applica- microarray data, proteomic data and population tions. The workshop will consist of a series of genetics. After genome sequencing, microarray invited talks and one or more sessions dedicated technology has emerged as a widely used platform to contributed posters. Please check periodically for genomic studies in the life sciences. DNA mi- the web-site http://3w.eco.uniroma1.it/bayes croarrays have high-density orderly arrangements to get updated information. For any fur- of nucleic acid spots. Many research studies have ther information please contact Luca Tardella demonstrated the general usefulness of genome ([email protected]). probing using microarrays. While simultaneous Scientific Committee: J.O. Berger (Duke Uni- measurement of thousands of gene expression lev- versity), P. Damien (University of Michigan), N.J. els provides a potential source of profound knowl- Hjort (University of Oslo), J.K. Ghosh (Indian Sta- edge, success of the microarray technology de- tistical Institute), S. Petrone (Universita` L. Bocconi), pends on the precision of the measurements and on G. Salinetti (Universita` di Roma ”La Sapienza”), the integration of computational tools for data min- Local Organizing Committee: P.Conti (Universita` ing, visualization, and statistical modeling. With di Roma ”La Sapienza”), M.Dall’Aglio (Universita` the abundance of data produced from genomic ”G.D’Annunzio” - Chieti - Pescara), B.Liseo (Uni- studies, the greatest challenge is analytical. The im- versita` di Roma ”La Sapienza”), L.Petrella (Univer- pact of genomic and proteomic technology on biol- sita` di Roma ”La Sapienza”), G.Salinetti (Univer- ogy will depend heavily on bioinformatic methods sita` di Roma ”La Sapienza”), L.Tardella (Universita` and statistical analysis. Sophisticated data-mining di Roma ”La Sapienza”). and analytical tools are needed to correlate data ob- tained from the arrays, to group them in a mean- ingful way, and to perform statistical analysis in WNAR/IMS Western Regional. First announce- order to investigate hypotheses of interest. Ex- ment. June 27-30, 2004, Albuquerque, New Mexico perimental design and statistical methods provide The 2004 WNAR/IMS Annual Meetings will be powerful analytical tools to biologists for the study held June 27-30 in Albuquerque, New Mexico. The of living systems. Through statistical analysis and University of New Mexico will host the Meetings the graphical display of clustering and classifica- and all the events will take place in the newly ren- tion results, microarray experiments allow biolo- ovated Student Union Building within the main gists to assimilate and explore the data in a natural campus of UNM. The program of the Meetings will and intuitive manner. include Invited Speakers, Short Courses, Student The conference will provide a forum for dis- Papers and Social Events. The Program Chairs are cussion on statistical and bioinformatic meth- Dave Higdon (Los Alamos) and Jason Fine (Wis- ods for analyzing genomic data. It will also consin). The Local Arrangement Chairs are Ed provide a bridge for smooth transition of sta- Bedrick (Univ. of New Mexico) and Gabriel Huerta tistical methods from the genomic era to the (Univ. of New Mexico). More information will be proteomic era. For more information, visit available at http://www.wnar.org http://www.amstat.org/chapters/boston/ genomic/Main.htm
15 ISBA Bulletin, December 2003 NEWS FROM THE WORLD
Second International IMS/ISBA Joint Meeting. Duke University, and Prof. Sylvia Richardson, Im- First Announcement. Bormio, Italy (Italian Alps), perial College London. ”MCMSki”: The Past, Present, and Future of Gibbs In addition, the members of the program com- Sampling, Wednesday, January 12 to Friday, January mittee (see below) are well on their way to assem- 14, 2005. bling an invited program that is as attractive as Following up on the success of the first joint the conference venue. In particular, invited ses- international meeting of IMS (Institute of Mathe- sions are anticipated on: molecular biology spatial matical Statistics) and ISBA (International Society and spatiotemporal methods, bioinformatics and for Bayesian Analysis) last summer in Isla Verde, genetics statistical, data mining, MCMC algorithms Puerto Rico, the second such joint meeting will be and software, and modern nonparametrics. held in Bormio, Italy (site of the world ski cham- Buses (exact times to be determined) will be pionships) next winter, January 12-14, 2005. The available to transport participants between the con- unifying theme of the conference is MCMC and its ference site and Malpensa Airport in Milan (approx impact in the (by then) 15 years since the appear- 3-3.5 hour ride). Limited financial support for the ance of the Gelfand and Smith paper. Since this is travel of junior investigators (¡ 5 years since PhD) is a joint meeting of two diverse organizations, talks also anticipated for those presenting in one of the on a wide variety of topics (both Bayesian and non- two poster sessions (Wed and Thurs). Bayesian) are anticipated. Further details are available on the initial confer- Our plan is to have 3 plenary speakers, each of ence flyer: http://www.biostat.umn.edu/∼brad/ whom will speak for 50 minutes at the beginning flyer.doc or from the official conference website: of each day. This will be followed by an invited http://eco.uninsubria.it/webdocenti/IMS-ISBA-05 session (3 speakers), then lunch, and then an af- as they become available. Please check the web ternoon break (where skiing and visiting the hot site regularly. In particular, more details on the spring spa will be among the options). Follow- registration fee and the accommodations will be ing the break will be another invited session, then available shortly. dinner and posters; in short, ”Valencia style” with Program Committee: Brad Carlin, University ski/spa time replacing the usual beach time. There of Minnesota (Co-Chair) Antonietta Mira, Univer- may also be ”satellite” events on Jan 11 or 15, such sity of Insubria (Co-Chair) Steve Brooks, Cam- as short courses or workshops on topics of particu- bridge University Montserrat Fuentes, North Car- lar interest. olina State University Paolo Giudici, University of The three plenary speakers will be: Prof. Persi Pavia, Giovanni Parmigiani, Johns Hopkins Uni- Diaconis, Stanford University, Prof. Alan Gelfand, versity.
ISBA/SBSSARCHIVEFOR ABSTRACTS All authors of statistics papers and speakers giving conference presentations with substantial Bayesian content should consider submitting an abstract of the paper or talk to the ISBA/SBSS Bayesian Abstract Archive. Links to e-prints are encouraged. To submit an abstract, or to search existing abstracts by author, title, or keywords, follow the instructions at the abstract’s web site, www.isds.duke.edu/isba-sbss/
16 Executive Committee Program Council Board Members
President: Ed George Nicky Best, Michael Goldstein, Ed- Past President: David Draper uardo Gutierrez-Pe´ na,˜ Pilar Iglesias, Chair: Alan Gelfand President Elect: Jim Berger Jun Liu, Anthony O’Hagan, Sonia Vice Chair: Jose´ Bernardo Treasurer: Peter Muller¨ Petrone, Raquel Prado, Christian Past Chair: Luis Pericchi Executive Secretary: § ¤ Robert, Fabrizio Ruggeri, Marina Web page: www.bayesian.org Cindy L. Christiansen ¦ ¥ Vannucci, Robert Wolpert EDITORIAL BOARD
Editor Interviews Brunero Liseo
Hedibert Lopes
17