The Era of “Big Data” Is Upon Us

Home , David Spiegelhalter, Julian Besag

THE ISBABULLETIN

Vol. 18 No. 2 June 2011 The oﬃcial bulletin of the International Society for Bayesian Analysis

A MESSAGE FROM THE PRESIDENT

THE ERAOF BIG DATA Big Data regime requires joint consideration of systems issues (how to store, index and transport data at massive scales; how to exploit parallel - Michael I. Jordan - and distributed platforms), statistical issues (how ISBA President, 2011 to cope with errors and biases of all kinds; how to [email protected] develop models and procedures that work when both n and p are astronomical), and algorithmic The era of “Big Data” is upon us. It is a rare issues (how to perform computations using re- week in which one does not see an article in a ma- sources that scale as linear or sub-linear functi- jor international magazine or newspaper on the ons of n and/or p). And of course there are legal, phenomenon of Big Data. One sees discussions commercial and social issues. of terabytes, petabytes, exabytes and zettabytes. One sees responses to these challenges at all le- (Megabytes and gigabytes seem to be viewed as vels of society. Companies are increasingly loo- old-fashioned.) king to hire people who have expertise in data Beyond discussions of mere size (and rates of analysis. . . . Continued on page 2. growth in size), the focus is increasingly on inferential issues. Past research on databases and information retrieval is viewed as having focused In this issue on mere storage, lookup and search, whereas the opportunity now is to discover new phenomena ➤ A MESSAGE FROM THE BA EDITOR ☛ and to “learn.” There is also an increasing awa- Page 3 reness that such opportunities are counterbalan- ➤ FROM THE PROGRAM COUNCIL ced by dangers of “discovering” phenomena that ☛Page 4 are not real. ➤ ISBA - SECTIONS Big Data issues arise both in science and tech- ☛Page 4 nology. In science, massive streams of data ha- ➤ ADRIAN F.M. SMITH Conference ve become the norm in areas such as astronomy, ☛Page 6 high-energy physics, ecology, genetics and mole- cular biology. In technology, there is a new drive ➤ ANNOTATED BIBLIOGRAPHY ☛ towards personalization, based on the increasing Page 7 availability of data on ﬁne-grained aspects of hu- ➤ SOFTWARE HIGHLIGHT man behavior; the analysis of such data is viewed ☛Page 11 as permitting the development of new services ➤ STUDENTS’ CORNER that are tailored to individuals. ☛Page 12 There seems to be a growing awareness that ➤ NEWS FROM THE WORLD the challenges involved in meeting Big Data pro- ☛Page 16 blems are interdisciplinary. Data analysis in the ISBA Bulletin, 18(2), June 2011 A MESSAGE FROM THE PRESIDENT

MESSAGE FROM THE PRESIDENT, Continued sciences, Big Data problems often arise in the from page 1. . . . While the job descriptions someti- context of “standard models,” which are often mes refer to “statistics,” they often refer to “da- already formulated in probabilistic terms. That ta analytics,” “data science,” “data mining” and is, significant prior knowledge is often present “machine learning.” In any case, there seems to and directly amenable to Bayesian inference. (3) be an increasing awareness that many of the fun- Consider a company wishing to offer personali- damental underlying problems are statistical in zed services to tens of millions of users. Large nature. Funding agencies are paying attention to amounts of data will have been collected for so- Big Data, and there is steady growth in the num- me users, but for most users there will be little bers of workshops and conferences devoted to or no data. Such situations cry out for Bayesian Big Data issues. hierarchical modeling. (4) The growing field of What is the role of Bayes in all of this? To da- Bayesian nonparametrics provides tools for de- te, I would venture to say “unfortunately rat- aling with situations in which phenomena conti- her little.” When discussions eventually turn to nue to emerge as data are collected. For example, methodology—in the job descriptions or the con- Bayesian nonparametrics not only provides pro- ference talks—one sees discussion of ideas such bability models that yield power-law distributi- as hierarchical clustering, K-means, the support ons, but it provides inferential machinery that in- vector machine, boosting, random projections, corporate these distributions. nearest neighbor and the Lasso. Rarely is Bayes But there remains the elephant in the room: mentioned. will the traditional tools of Bayesian posterior in- I am occasionally invited to give talks in work- ference be able to cope with arbitrarily large data shops on Big Data, and my somewhat contrarian streams? To this I would respond that we simply nature leads me to give talks on Bayesian approa- do not know. Although one can certainly imagi- ches. This is met with immediate skepticism— ne that algorithms such as MCMC and sequenti- are you telling us that we should be rolling al Monte Carlo will scale increasingly poorly as out MCMC methods in the Big Data world, n and p increase without bound, it is also con- when such methods are often too slow for old- ceivable that this will not happen in general. Per- fashioned small-data problems? haps the sharing of statistical strength implemen- I am skeptical about this skepticism, but I ted by Bayesian hierarchical models will lead to am rarely able to formulate a very convincing implicit computational efficiencies, at least for so- response. Unfortunately, the Bayesian literature me aspects of the posterior distribution (ideally, has not yet taken Big Data very seriously, and we for the inferences that matter). There are research simply do not know what kinds of Bayesian me- questions here that have not begun to be touched. thodology might be viable at massive scales. The recent growth of activity in parallel and Another criticism goes as follows: Bayes is distributed implementations of posterior infe- principally concerned with error bars rather than rence algorithms (e.g., GPUs) is encouraging in point estimates, and in the realm of Big Data this regard. But we should not view parallelism the error bars are negligible, so why do we need as a silver bullet; it will not suffice to scale Baye- Bayes? This argument is easier to rebut. In analy- sian inference to massive data sets. We should sis of Big Data one is rarely concerned with quan- instead hope that this line of research will lead tities such as population means; rather one is to increased understanding of how posterior in- concerned with small sub-populations, and conference algorithms behave at larger scales, hope- cerned with the tails of distributions. Moreover, fully pointing the way towards further modeling Big Data analysis generally involves many error- and algorithmic developments. prone steps, and the propagation and control of Finally, I wish to make a somewhat philoso- uncertainty is critical. phical note, one which is as applicable to non- In general I believe that Bayes has several na- Bayesian statistics as to Bayesian statistics. Sup- tural advantages in the world of Big Data. Let pose that I am a statistician faced with an in- me note four such advantages: (1) Analyses of ference problem involving a possibly large data Big Data often have an exploratory flavor rather stream. Sallying forth with a fast computer and than a confirmatory flavor. Some of the concerns the R package under my arm, I am pleased to over family-wise error rates that bedevil classi- be given more and more data, up to a certain cal approaches to exploratory data analysis are point, beyond which I start to become concer- mitigated in the Bayesian framework. (2) In the ned. I eventually throw up my hands, unable to

Content 2 www.bayesian.org ISBA Bulletin, 18(2), June 2011 BAYESIAN ANALYSIS - A MESSAGE FROM THE EDITOR do much of anything. This seems odd in a cer- which must be taken into account in the budget.) tain philosophical sense. If data are our princi- Probably the answer is yes, but we are very far in pal resource, why should having too much data the current state of our discipline from being able cause such embarrassment? Shouldn’t I be able to provide such guarantees. to “throw away” or “reduce” data as they accrue In short, I would propose that we begin to and still maintain a certain quality of statistical think about how to do Bayesian data analysis in inference within a given computational budget? the context of data sets of arbitrary size; let’s aim Indeed, despite my limited computational bud- to overcome our embarrassment at the riches of get, shouldn’t I achieve a growing quality of sta- massive data once and for all. tistical inference as data accrue? (Note that the act The Big Data problem is here to stay, and Baye- of data reduction involves a computational cost, sians should begin to think Big. ▲

AMESSAGEFROMTHE EDITOR cipation to those who may be interested in these areas of research. As every quarter, you will also ﬁnd the usual Manuel Mendoza sections of the Bulletin. In particular, I call your [email protected] attention to the Annotated Bibliography Section where the Editor, Beatrix Jones, asked Prof. An- In this Bulletin our President, Michael Jordan, thony Pettitt to write an article on the main con- calls our attention on the increasing number of tributions of the late Julian Besag. The article in- problems where the data, far from being scarse, is cludes an impressive list of areas where Prof. Be- massive. The need for statistical procedures, ca- sag contributed decisively. Our Software Editor, pable to deal with huge amounts of complex data Alex Lewin, invited Jong Hee Park to introdu- to produce useful, effective and timely results is ce some recent developments of the R package discussed in detail. MCMCpack. In addition, our Student’s Corner Also in this issue, we have a report by Paul De- Editor, Luke Bornn, continues his quest and po- mian on the Conference that was recently held ses one more question to his panel of distiguished in Creete, in honour to Prof. Adrian F.M. Smith. colleagues. Again, the result is interesting and re- More than a hundred colleagues, many of them vealing. coauthors and/or former students of Sir Adrian Finally, I want to encourage all members of IS- Smith, gathered in the Aegean Sea in an impres- BA to contribute to the Bulletin with their sugge- sive academic tribute to one of our leaders. stions, manuscripts and announcements. Please Brief notes from the sections of ISBA describe do not hesitate to contact me or any member of their recent activities and call for an active parti- the Editorial Board. ▲

BAYESIAN ANALYSIS - A MESSAGE FROM THE EDITOR

UPDATE FROM BA tributions. We also are welcoming in Kary Myers, who is our new Production Editor, and we are ex- Herbie Lee cited to have her join our team. Editor-in-Chief The June issue of BA features a paper on cova- [email protected] riance structures for multidimensional arrays by The big event at Bayesian Analysis this quar- Peter Hoff. Use of the Tucker product provides ter is that we are saying goodbye to Angelika a practical framework for approaching the pro- van der Linde, who is stepping down after ma- blem. Additional discussion and perspective ap- ny years as our Production Editor. Angelika has pears in discussions by Genevera Allen and He- been crucial in helping us produce quality issues dibert Lopes. This issue also contains other ﬁne and we need to thank her for her efforts and con- articles in computational statistics, bioinforma- tics, stochastic processes, and nonparametrics.▲ Content 3 www.bayesian.org ISBA Bulletin, 18(2), June 2011 FROM THE PROGRAM COUNCIL

FROM THE PROGRAM COUNCIL

ISBANEWS (session abstract), and the list of 4 speakers and should be sent to [email protected] by Igor Pruenster September 18th, 2011. [email protected] The ISBA 2012 World Meeting – the premier conference of the International Society for Bayesi- Call for special topic sessions at ISBA 2012 an Analysis (ISBA) – will be held in Kyoto, Japan, from June 25 to June 29, 2012. Preliminary pro- The Program Committee of ISBA 2012 invites gram can be found at http://www2.e.u-tokyo. submissions of proposals for special topic sessi- ac.jp/∼isba2012/. ons. Each session consists of 4 talks of 25 min Details about support for junior presenters will with a common theme. The proposal should in- be ﬁnalized in late 2011. Please check the confe- clude the title of the session, a brief description rence website for updates.▲

ISBA - SECTIONS

OBJECTIVE BAYESIAN in ECNU’s 60th anniversary celebration, appro- SECTION priate as the university started the ﬁrst mathematical statistics department in China. ECNU al- Jim Berger so used the occasion to inaugurate a new Aca- Chair demy of Applied Statistical Sciences, an acade- [email protected] my in which Dongchu Sun (Program Chair of the OBayes section) will play a central role. OBayes 2011. This workshop, formally the Since students in China typically receive only 2011 International Workshop on Objective Bayes modest exposure to Bayesian analysis, the con- Methodology, was held in June 11-15, 2011 at ference began with a full day of tutorials. There East China Normal University in Shanghai, Chi- were then a total of 21 invited talks with discus- na. The conference had roughly 100 participants, sants, followed by lively ﬂoor discussion; many half from China and half from other countries. of the talks are posted at http://www.sfs.ecnu. It was designated as one of the principle events

Content 4 www.bayesian.org ISBA Bulletin, 18(2), June 2011 ISBA - SECTIONS edu.cn/Obayes2011/index.html. Other confe- The program committee and local organizers rence highlights included a well-attended poster of the workshop - led by Dongchu Sun and Yin- session with the requisite refreshments (best stu- cai Tang - did a wonderful job, and ISBA played dent poster winners were Xiaojing Wang from a valuable role in handling the registration and Duke U., Anchu Xu from ECNU, and Chang promoting the workshop. Xu, U. Missouri); wonderful meals, including an This was the ﬁrst formal meeting of the Objec- awesome banquet with hilarious after dinner re- tive Bayesian Section of ISBA and, I believe, was marks by Ed George; and a spectacular tour of the ﬁrst Bayesian meeting ever held in China; we Shanghai - including a delightful boat cruise that hope there are many more to come! had no problems whatsoever; perhaps the “Baye- OBayes 2013. Please send me any suggestions sian conference boat excursion curse” has been you might have concerning the location and ti- lifted! ming of the 2013 OBayes meeting.▲

NONPARAMETRIC BAYES ciation (AME) and the International Society for SECTION Bayesian Analysis (ISBA). Here, we had the opportunity to hear 19 invi- Ramses Mena ted presentations as well as 36 contributed talks Program Chair together with 4 tutorials and to enjoy a superb [email protected] view of the portuary activities from the wonderful lounge located at the top ﬂoor of the Precisely one year ago, in June 2010, Stephen hotel where the posters session was organized. Walker wrote a brief article for the Bulletin an- Around 150 well stablished researchers, young nouncing the Nonparametric Bayes Section of IS- scientists and students from 15 different coun- BA. Since then, the nonparametric community tries and 67 institutions attended the workshop has been working hard and while this note is be- in a warm and friendly environment. en written, the 8th Workshop on Bayesian Non- Finally, I would like to thank all the people, parametrics is taking place in Veracruz, Mexico. institutions and societies who support us whi- The objective of the workshop is to bring to- le organizing the workshop and as for the sec- gether experts and young talented scientists de- tion, please send me any comments, suggestions voted to the study and application of Bayesian or questions you might have regarding the acti- nonparametric techniques. It was organized un- vities of the section. It will be great to hear from der the auspices of the Mexican Statistical Asso- you.▲

Content 5 www.bayesian.org ISBA Bulletin, 18(2), June 2011 ADRIAN F.M. SMITH Conference

ADRIAN F.M. SMITH Conference

HIERARCHICAL MODELSAND do Gutierrez-Pena, prompting Adrian to quip, MARKOV CHAIN MONTE CARLO ”how many Mexicans does it take. . . ”. The ban- JUNE 2 TO 5, 2011. CREETE,GREECE. quet dinner was held by the main swimming pool. The highlights of the after-dinner show in- Paul Damien cluded a humorous Monty Python skit by Linda [email protected] Sharples, Jon Forster, Nick Polson, and Peter Mu- eller. The inimitable Ed George proceeded next Set against the back drop of sunny skies, cool with a series of (by his own admission) ´´tasteless breezes, and the bluish green waters of the Cre- jokes”, and concluded by poignantly noting that tan Sea, the Knossos Royal Village was the ve- ”we all miss Adrian”. A brief slide show of Adri- nue for the conference honoring Adrian Smith, a an’s photographs, set to the tunes of the Irish Bayesian statistician whose many path-breaking tenor John McCormack, was then followed by a contributions to the theory and practice of stati- roasting of Adrian by that natural-born (Bayesi- stics continue to impact academic research and an) stand-up comedian John Deely. Finally, the industrial practice world-wide. moment that we had all been waiting for: Adri- an took the podium. His flawless and inspiring Conference Facts 5-minute delivery confirmed, once again, why he holds a singular position in the advancement Approximately 125 to 130 people attended the of statistical science. The entire show was filmed event, which excluded parallel sessions, thus for- by Wes Johnson. The film will be posted on the cing Adrian to participate in all the talks! There website www.afmsmith.com in July. were a total of 32 talks, each 30 minutes long, and 17 poster presentations. The speakers we- The Talks re encouraged to address broad statistical ideas, focusing, if possible, on the twin themes of Hier- Applying a valuable Adrian Smith dictum, na- archical Models and Markov Chain Monte Carlo mely the importance of brevity in communicati- -the banner heading for the conference. To ensure on, the welcoming remark was made by Petros a leisurely pace, the sessions were interspersed Dellaportas, who kicked off the event by stating, with multiple coffee breaks, culminating in an ”Welcome”. He then introduced the first spea- early afternoon siesta at 15:30 hours, followed ker David Spiegelhalter. David’s aptly titled talk, by dinner at 20:00 hours. On the final day, af- ´´Following the leader of the revolution”, was ter the last lecture, several group photos were wildly entertaining and scientifically informati- taken by Manuel Mendoza-Ramirez and Eduar- ve; it set the tone for the rest of the meeting. The

Content 6 www.bayesian.org ISBA Bulletin, 18(2), June 2011 ANNOTATED BIBLIOGRAPHY

ﬁnal speaker was Gareth Roberts. He bookended David’s talk by punctuating his own with a se- The Organizing Committee ries of laugh-out-loud anecdotes and jokes, while cleverly interjecting them with deep insights on Paul Damien, Petros Dellaportas, Jon Forster, convergence issues in MCMC within a hierarchi- Nick Polson, Gareth Roberts, David Stephens, cal modeling framework. The speakers list may Jon Wakeﬁeld▲ be accessed at www.afmsmith.com.

ANNOTATED BIBLIOGRAPHY

THE CONTRIBUTIONSOF JULIAN from Nottingham or Loughborough (where Juli- BESAG an’s father had been a lecturer). Later, I came face to face with him over an interview panel in 1982 James McKeone & Tony Pettitt for a lecturing job at Durham. Julian asked me [email protected] why I had not had a paper read to the RSS and, [email protected] somewhat tritely, I replied that I had been in Au- stralia for two years. I do not think the answer was well received. This article is joint work with James McKeone. James is In 1991 I had a position in Australia at QUT. just starting a PhD at QUT which involves a problem from I organized a statistical computing and biome- neurology and the modelling of action potentials recorded trics meeting at Coolangatta on the beach on the at different stimulus levels on an 8 by 16 regular array of Queensland / New South Wales border. I had in- 128 electrodes. There are some interesting spatial data with vited Sandy Weisberg and Julian as the statistical considerable structure. A Bayesian approach is planned, and computing invited international speakers. Julian James enjoyed very much reading several of Julian’s papers. gave a talk on Gibbs sampling and spatial stati- Tony Pettitt stics and I remember his carefully drawn over- head transparencies of spatial DAGs. That must have been the first introduction in a public lec- News of Julian’s death was very sad but ga- ture of the those ideas to an Australian audience. ve me the opportunity to reflect on his immense It sowed some Bayesian seeds not least with my contribution to modern statistics and my perso- colleague at QUT, Kerrie Mengersen. nal contact with him. In the remarks below the two read papers of I must have first met Julian at one of the 1974 and 1986 are given prominence. “Spatial in- monthly RSS read paper meetings which I atten- teraction and the statistical analysis of lattice sy- ded regularly, getting the train down to London stems”(1974) was considered seminal at the time

Content 7 www.bayesian.org ISBA Bulletin, 18(2), June 2011 ANNOTATED BIBLIOGRAPHY it was published and more than lived up to this Professor D. R. Cox (Imperial College): The pa- promise. I was present at the 1986 meeting where per is original, lucid and comprehensive. The to- “Statistical analysis of Dirty pictures”was read. pic is important and notoriously difficult. It is a At the end of Julian’s presentation he received a pleasure to congratulate Mr Besag. great acclamation, even though we never saw the Professor P. Whittle (University of Cambridge): “dirty pictures” or perhaps relieved that we had Mr Besag has given us a substantial, inte- not. He gave the talk with the confidence of so- resting and most useful paper. It collects to- meone who knew absolutely that he had a win- gether a great amount of material and clarifies ner (he was the most highly cited mathematical a number of points admirably. For example, the scientist of the 1980s). And that was always the Hammersley-Clifford theorem now appears ve- way Julian gave a seminar, with immaculate pre- ry much more accessible. paration and confident delivery. Dr J. M. Hammersley (University of Oxford): I am much impressed by Mr Besag’s very intere- Conditional Framework sting and far-ranging paper on spatial interaction. The variety and flexibility of his treatment • J. E. Besag (1972) “Nearest-neighbour sy- will greatly contribute to our understanding of stems and the auto-logistic model for bi- these very difficult problems of co-operative phe- nary data.” Journal of the Royal Statisti- nomena. cal Society Series B (Methodological) 34:75– Professor M. S. Bartlett(University of Oxford): 83. Considers a conditional framework for I regret that my absence in Australia has pre- spatially dependent random variables on vented my coming to this meeting, but it is in- a regular lattice. Finding that in the case deed a pleasure to send my congratulations to of binary homogeneous random variables, the author for this very comprehensive and va- the only theoretically plausible model is luable paper on nearest-neighbour lattice mo- the now familiar auto-logistic process. So- dels, containing both exposition of basic theory me spatio-temporal models are investiga- and of statistical techniques of analysing spatial ted but they are noted to be of limited prac- interaction. I find his discussion of auto-normal tical appeal. and auto-logistic model-fitting especially interesting and useful. • J. E. Besag (1974) “Spatial interaction These comments indicate the substantial and the statistical analysis of lattice sy- standing and seminal nature of the paper. stems.”Journal of the Royal Statistical So- The discussion points to the growing awa- ciety Series B (Methodological) 36:192–236. reness of graphical modelling as a tool in Further develops the ideas seeded in Besag statistics. (1972). The focus of the paper is on binary and Gaussian models with applications • J. E. Besag (1977). “Errors-in-variables esti- in plant ecology. A simple, alternate proof mation for Gaussian lattice schemes.” Jour- of the Hammersley-Clifford theorem is pre- nal of the Royal Statistical Society. Series B sented assuming the positivity assumption. (Methodological), 39:73–78. Some notes on The theorem may be used to specify speci- estimation when the variates of primary in- fic spatial schemes under modest assumpti- terest are not observable on a simple Gaus- ons. Methods for inference are also presen- sian lattice. Notes on estimation procedure ted along with a coding technique for mo- for error-in-variables approximation. del fitting that is easy to implement and has a sound theoretical basis, giving a true but inefficient likelihood. The paper continues Extensions including Pseudoli- to argue the merits of modelling under the conditional framework as opposed to the kelihood alternative joint framework. • J. E. Besag (1975) “Statistical analysis of The paper was a read paper at the RSS and non-lattice data.” The Statistician 24:179– these are some of the discussants’ publis- 195. Introduces the pseudolikelihood ap- hed comments. proach to estimation of spatial interaction of random variables for irregularly dis-

Content 8 www.bayesian.org ISBA Bulletin, 18(2), June 2011 ANNOTATED BIBLIOGRAPHY

tributed data points. The method permits test that performs well in every circumstan- more computationally tractable parameter ce - rather a general approach is preferred, estimation by computing an approximation that even when analytical (asymptotic) dis- to the likelihood for a set of random varia- tributional theory is known, an MC test is bles. often also worthwhile, and that the nature of MC tests is fairly intuitive and easy to • J. E. Besag and P. A. P. Moran (1975). “On communicate to non-statisticians. the estimation and testing of spatial interaction in Gaussian lattice processes.” Bio- metrika 62:555-562. Develops a technique Bayesian spatial modellling and for explicit maximum likelihood estimation for first-order, regular lattice autonor- MCMC mal models. The estimation is computed on an infinite lattice scheme, and on a torus • J. E. Besag (1986). “On the statistical ana- lattice with first-order, isotropic square de- lysis of dirty pictures.” Journal of the Roy- pendence structure. The coding technique al Statistical Society. Series B (Methodologi- is found to be efficient only for relatively cal), 48(3):259–302. Assumes the local cha- small first order correlation. racteristics of the true image are modelled by a non-degenerate Markov random field. • J. E. Besag (1977). “Efficiency of pseudo- Uses a priori information in a Bayesian fra- likelihood estimation for simple Gaussian mework to determine the maximum a po- fields.” Biometrika 64:616–618. Adjunct to steriori (MAP) estimate giving the colouring Besag (1975) and Besag and Moran (1975). that possesses overall maximum probabi- Notes on efficiency of pseudolikelihood lity given the data. Introduces the iterated estimation contrasted to the maximum like- conditional modes (ICM) algorithm to aid lihood approach of the former papers. Be- in parameter estimation due to the compu- sag notes that the psuedolikelihood techni- tational difficulty of the new model. The que produces “encouraging” efficiency re- ICM algorithm uses simulated annealing sults for the first order isotropic regular lat- and a Gibbs like sweep though pixels fin- tice schemes. ding the posterior most likely value for the pixel. Examples include some noisy images involving true scenes with a number of co- Monte Carlo based inference lours and show substantial improvements over joint pixel maximum likelihood classi- • J. E. Besag and P. Clifford (1991). “Se- fiers. quential Monte Carlo p-values.” Biometri- ka 78:301–304. The computational burden For computationally efficient estimation of of the using Monte Carlo methods to as- parameters involved in the Markov ran- sess statistical significance is notably high. dom field, the paper recommends maxi- This paper presents a method of sequential mum pseudo-likelihood. sampling where samples are drawn until a This paper was a discussion paper at a RSS predetermined number of deviates exceed meeting and the following comments are of a benchmark value. The ratio of the num- historical interest. The first was most pro- ber of deviates that exceed the benchmark phetic in terms of impact of the paper and to the total number simulated is an estima- the second shows the seeds of BUGS were tion to the true p-value that is exact when germinating. H0 is true and provides a good approximation when H is false. 0 Professor D. M. Titterington (University of • J. E. Besag and P. J. Diggle (1977). “Sim- Glasgow): A crucial feature of tonight’s paper ple Monte Carlo tests for spatial pattern.” is the proposal of a joint probability function ... Journal of the Royal Statistical Society, Se- To put it bluntly, it could be said that the sty- ries C (Applied Statistics), 26:327–333. A re- lish modelling turns out to be a means to the view of Monte Carlo tests for spatial pat- end of proposing, in the form of the ICM me- tern with an emphasis on preliminary inve- thod, a practically feasible, if rather ad hoc algo- stigation. Concludes that there is no specific rithm ... and Professor Besag is to be applauded

Content 9 www.bayesian.org ISBA Bulletin, 18(2), June 2011 ANNOTATED BIBLIOGRAPHY

for leading UK statisticians into this important Algebra of interacting Systems area. Maybe his technique’s title can be exten- ded eponymously to that of Iterated Conditional • J. E. Besag (1981). “On a system of two- Besag Modes to reflect, by its initials, his inter- dimensional recurrence equations.” Journal national and explosive impact on the field! (a of the Royal Statistical Society. Series B (Me- contemporary reference to the state of the thodological), 43:302–309. Considers two- cold war). dimensional recurrence equations of a first order auto-normal scheme on a first-order, Dr D. J. Spiegelhalter (MRC Biostatistics Unit, isotropic, doubly infinite rectangular latti- Cambridge): A number of aspects in tonight’s ce. It is shown that the covariance functi- stimulating paper have parallels in current on of a Gaussian Markov random field can work on expert systems. Thus in place of the be approximately represented in terms of pixels in image analysis, we substitute possible a modified Bessel function that decreases clinical findings and diseases, while the regu- monotonically with distance. lar lattice structure of the relationships expressed in the random field is replaced by a complex • J. E. Besag and C. Kooperberg (1995). “On network comprising the local relationships ar- conditional and intrinsic autoregression.” chetypically expressed as subjective ’rules’. ... Biometrika, 82(4):733–746. A summary of Gaussian conditional autoregressions - int- • J. E. Besag, J. York, and A. Mollie´ (1991). rinsic or otherwise. Illustrates how Demp- ”Bayesian image restoration, with two app- ster’s algorithm may be used to alleviate lications in spatial statistics.” Annals of the the problem of an undesirable marginal va- Institute of Statistical Mathematics, 43:1–59. riance structure in CAR models. Shows that Interprets spatial data analysis problems as under the intrinsic CAR model, the para- problems in image analysis. Introduction of meters may be estimated such that Xi and the conditional autoregression (CAR) mo- Xj have constant variance when i and j are del and a rank deficient intrinsic alternate neighbours. (ICAR or IAR). The methodology empha- • F. Bartolucci and J. E. Besag (2002). “A sizes a Bayesian approach, using a Gibbs recursive algorithm for Markov random sampler to compute the maximum of the fields.” Biometrika, 89(3):724–730. A recur- posterior marginal probabilities. Much ci- sive estimation procedure for the full con- ted in the disease mapping literature. ditionals of MRFs as an alternative to the Brooks expansion. The technique uses ideas • J. E. Besag and P. J. Green (1993). “Spatial from time series factorisation and a block statistics and Bayesian computation.” Jour- Gibbs sampler. As with Besag (1974), the nal of the Royal Statistical Society. Series B factorisation is arbitrary and alternate fac- (Methodological), 55:25–37. Presents deve- torisations must lead to the same joint dis- lopment of and a review of MCMC techni- tribution. ques in Bayesian inference. The techniques that are discussed are variance reduction • J. E. Besag and D. Mondal (2005). “First- methods, antithetic variable methods, con- order intrinsic autoregressions and the De vergence and efficiency, use of auxiliary va- Wijs process.” Biometrika, 92:909–920 De- riables, and MCMC under multimodality. velops relationships between first-order in- With some application to field experiments. trinsic autoregressions and the de Wijs process which involves a continuous space dif- ferential equation. • J. E. Besag, P. Green, D. Higdon, and K. Mengersen (1995). “Bayesian computation and stochastic systems.” Statisti- More applied cal Science, 10:3–66. Basic methodology of MCMC with an emphasis on the Bayesi- • P. J. Diggle, J. E. Besag, and T. Gleaves an paradigm and modelling spatially de- (1976). “Analysis of spatial point patterns pendent systems. Applications in spatial or by means of distance methods.” Biometrics, temporal aspects of the problems, agricul- 32(3):659–667. Considers some distance ba- tural field experiments, image analysis. sed method to test for spatial randomness.

Content 10 www.bayesian.org ISBA Bulletin, 18(2), June 2011 SOFTWARE HIGHLIGHT

The emphasis is on their use at the preli- • L. Knorr-Held and J. Besag (1998). “Model- minary stage of the investigation. T-square ling risk from a disease in time and space. sampling is promoted as a quick and useful Statistics in Medicine.” 17:2045–2060, 1998. method that can be used between different Develops ideas for modelling risk from a populations and is suited to large populati- disease in time and space using the basic ons. models introduced for image analysis.▲

SOFTWARE HIGHLIGHT

BAYESIAN TIME SERIES MODELS on model with multiple parameter breaks using IN MCMCPACK the Cowles method [3]. The simulation of latent states is based on Park (2011) [8]’s method to Jong Hee Park avoid the problem of clustered ordered catego- [email protected] ries. Last, MCMCpoissonChange fits a Poisson regression model with multiple changepoints. To avoid computational difficulties in designing the Bayesian techniques have been widely app- proposal density for Metropolis-Hasting algo- lied in social science research, and there is litt- rithms, MCMCpoissonChange utilizes Fruhwirth- le doubt that R packages for Bayesian inference Schnatter and Wagner (2006) [4]’s auxiliary mix- [9] continue to play an important role in these ture sampling method. The details of the mo- applications. Among them, MCMCpack [6] sup- del are discussed in Park (2010) [7]. For each of ports Bayesian analysis for a wide range of sta- these change-point models, MCMCpack allows tistical models, density functions, and pseudo- Bayesian model comparison using marginal like- random number generators, as well as providing lihoods [1]. Thus, users can identify the “true” a number of utility functions for creating graphs, number of breaks by comparing several candida- reading and writing data to external files, create models with varying numbers of breaks before ting mcmc objects, and manipulating variance- proceeding to the interpretation of the posterior covariance matrices. One characteristic that di- distributions of the “true” model. stinguishes MCMCpack from other Bayesian R packages is its integrated Bayesian approach to In addition to the above-described change- model-fitting and model diagnostics. In this ar- point models, MCMCpack also includes dy- ticle, I introduce some recent developments in namic models for two specialized social MCMCpack, with a particular focus on Bayesian science model applications. First is the one- time series models. dimensional item response theory (IRT) model. MCMCdynamicIRT1d simulates from the posteri- There are two kinds of time series models in or distribution using the algorithm of Martin MCMCpack. First, there are change-point models and Quinn (2002) [5]. The model assumes that (Markov mixture models or hidden Markov mo- each subject has a given ability and that each dels) using discrete response data. MCMCpack item has both a difficulty parameter and a dis- uses Chib [2]’s efficient algorithm to sample hid- crimination parameter. The observed choice is den states from a variety of parametric models. dictated by an unobserved utility which is a For binary time series data, MCMCbinaryChange function of item-specific characteristics, subject- and MCMCprobitChange provide functions to find specific abilities, and disturbances. The second breaks and estimate regime-specific parame- specialized dynamic model in MCMCpack is ters given detected breaks. While regressors are MCMCdynamicEI, which fits Quinn [9]’s dynamic not allowed in MCMCbinaryChange, users may ecological inference model for partially observed employ the same modeling structure with a 2 x 2 contingency tables. MCMCdynamicEI incor- probit regression model in MCMCprobitChange. porates a dynamic random walk prior into Wa- MCMCoprobitChange simulates from the poste- kefield (2001)[11]’s model of ecological inference rior distribution of an ordinal probit regressi- to capture temporal dependence in 2 x 2 contin-

Content 11 www.bayesian.org ISBA Bulletin, 18(2), June 2011 STUDENTS’ CORNER gency tables observed over time. [4] S. Fruhwirth-Schnatter, Wagner H. Auxiliary Mixture Sampling for Parameter-driven Models The functions included in MCMCpack have in- of Time Series of Small Counts with Applications creased dramatically in recent years. However, to State Space Modelling. Biometrika, 93, 827– thanks to its unified design, all of these functi- 841, 2006. ons share a standardized model interface, even as they cover a variety of model-specific, computa- [5] A. D. Martin, K. M. Quinn. Dynamic Ideal tionally efficient MCMC algorithms. MCMCpack Point Estimation via Markov Chain Monte Carlo continues to serve as an easy-to-use and highly for the U.S. Supreme Court, 1953–1999. Political reliable computational environment for a varie- Analysis, 10, 134–153, 2002. ty of statistical models in the social sciences. We, the developers of MCMCpack, welcome all com- [6] A. D. Martin, K. M. Quinn, J. H. Park. ments and suggestions from users. MCMCpack: Markov Chain Monte Carlo (MCMC) Package. R package version 1.0- 11, URL http://CRAN.R-project.org/ packa- Acknowledgements ge=MCMCpack. [7] J. H. Park. Structural Change in U.S. Presi- I thank the two other members of the the MCM- dentsO˜ Use of Force. American Journal of Poli- Cpack project, Andrew D. Martin and Kevin M. tical Science, 54(3), 766–782, 2010. Quinn. I would also like to thank all the users of and contributors to MCMCpack including Mi- [8] J. H. Park. Changepoint Analysis of Binary chael Malecki, Daniel Pemstein, and Craig Reed. and Ordinal Probit Models: An Application to Bank Rate Policy Under the Interwar Gold Stan- dard Political Analysis, Forthcoming, 2011.

References [9] J. H. Park, A. D. Martin, K. M. Quinn. CRAN Task View: Bayesian Inference. [1] S. Chib. Marginal Likelihood From the Gibbs Version 2011-05-10, URL http://CRAN.R- Output. Journal of the American Statistical Asso- project.org/view=Bayesian. ciation, 90, 1313–1321, 1995. [10] K. Quinn. Ecological Inference in the Pre- [2] S. Chib. Estimation and Comparison of Mul- sence of Temporal Dependence. In Ecological In- tiple Change-Point Models. Journal of Econome- ference: New Methodological Strategies. Gary trics, 86(2), 221–241, 1998. King, Ori Rosen, and Martin A. Tanner (eds.). New York: Cambridge University Press, 2004. [3] M. K. Cowles. Accelerating Monte Carlo Mar- kov Chain Convergence for Cumulative-Link [11] J. C. Wakeﬁeld. Ecological Inference for 2 x Generalized Linear Models. Statistics and Com- 2 Tables. Journal of the Royal Statistical Society, puting, 6, 101–110, 1996. Series A. 167(3): 385–445, 2004.▲

STUDENTS’ CORNER

Luke Bornn email me. Following the Q & A, ﬁnd the disser- [email protected] tation abstract of Minghui Shi, entitled “Bayesian Sparse Learning for High Dimensional Data.” If In this issue’s Students’ Corner, we continue you are newly graduated and would like to pu- our Q & A with a panel of leading Bayesian stati- blish your thesis abstract, don’t hesitate to con- sticians, including new member Paul Gustafson tact me. (University of British Columbia). If you have a question for the panel for future issues, please

Content 12 www.bayesian.org ISBA Bulletin, 18(2), June 2011 STUDENTS’ CORNER

Q&A pends not only on your past accomplishments, but also on your plans and their viability. The “In what circumstances would different models suits different people. There are you recommend your students not many fellowships, but a factor in choosing an do a post-doc? In what assistantship is the degree of independence you circumstances is a post-doc not will be allowed. Good luck! the best choice for graduates?” Paul Gustafson Peter Green [email protected] [email protected] My gut reaction is that doing a post-doc, whether by choice or necessity, is a good thing. So- This is another of those “it depends where you me time to focus on science, and transition from are” questions that Luke is so good at ... in this ca- doctoral student to independent investigator wise, because the availability of post-doc positions thout too many distractions, is a real opportunity varies from country to country, as do the implica- to start a scientific career on a sound footing. This tions for what happens next in your career. But in opinion is based both on my own experience of general, I would say “almost always” I would re- doing a one-year postdoc (albeit with some tea- commend a graduating student to do a post-doc. ching added-in), and from what I’ve seen with However, you should do it for the right rea- others launching their careers. To facilitate the sons. It is not about “buying time” to allow you to transition to independent investigator, I think a postpone a decision about your future, although good postdoc supervisor is commensurately mo- this can of course be a useful by-product. You re hands-off than a Ph.D. supervisor, but more should think of it as a transition and an oppor- hands-on than a research collaborator. tunity ... a transition between the relative securi- Of course the feasibility of instant transition, ty of being a graduate student, where duties are straight from Ph.D. to tenure-track, varies not few and responsibilities for planning research are just with the graduate’s CV, but also over time, shared with your adviser, and the independence as economic and demographic forces exert them- and greater demands of many kinds that imme- selves. As well, national boundaries come into diately hit you in a faculty position. As a post- play. For instance, the Canadian granting system doc, you take a step in these directions while re- is currently such that research funding levels in taining support and mentoring, and without ma- the statistical sciences are too low to create ve- king long-term commitments. As an opportunity, ry many postdoctoral positions, at least for pure- it allows a period to try new directions in rese- ly methodological work. More interdisciplinary arch, exploiting the freedom that comes from re- work may fare better. lease from the straitjacket that many students feel As a supervisor, I’ve seen both an outgoing constrains them, completing a thesis project to a Ph.D. graduate and an incoming postdoc get the deadline. best of both worlds. Both accepted a tenure-track I think the post-doc experience works best and position straight out of the Ph.D, but then nego- adds most if you take a step outside your comfort tiated to start one year later, in order to do a post- zone – consider a research project that is adven- doc year elsewhere in the meantime. This year turous, and not perfectly aligned to your current is then particularly unencumbered and research- experience and skills. And move to a different de- centric, with neither professorial duties or job- partment. Too many post-docs who don’t move search duties at hand. Someone sufficiently ta- spend their time with the same adviser and end lented and fortunate to realize this situation is up just doing the research they didn’t have time giving themselves a particularly good opportu- for in their thesis. nity to launch. Here in the UK, post-docs come in two flavours – research assistantships and research fellowships. In the first, the PI has been awarded Dave Higdon funds to complete a specific project, and you are [email protected] the human resource to help him or her complete it; in the second, you decide the project and A post-doc position is a great opportunity to aim to be largely self-directed – appointment de- focus on an application, work with new people,

Content 13 www.bayesian.org ISBA Bulletin, 18(2), June 2011 STUDENTS’ CORNER

publish in a new area, see someplace new and ex- lated papers (and start new ones!), and get a good pand your experience, with only a two or three glimpse at academic and research life from the in- year commitment. The downside is that you’ve sider’s viewpoint. It is the opportunity of seeing got a few more years of apartment living, and oneself as a scholar. This is also the last opportu- you’ll likely have to move again when it’s over. I nity to enjoy of lots of free time to focus on re- know people who have done very well with and search work. Aside from these motivations, ma- without taking a post-doc job – I doubt the post- ny academic positions require de facto a recom- doc was the deciding factor in their success. mendation letter originating in a previous post- From a professional standpoint, postdoc expe- doctoral experience. rience can be a big help. At Los Alamos we are When there is no motivation for a research or always looking for people with experience and academic career, a post-doc could still be a good breadth, adept at connecting statistics to appli- way to acquire experience in some particular area cations. In addition, a postdoc position affords of application. But if the student only wants a re- the opportunity to build some expertise in a new gular job or a teaching position, then the post-doc application area. This is particularly important if does not seem recommendable as it would only your research is going to be guided by applicati- delay the final goal. ons. Of course, non-professional aspects of a new opportunity can sometimes outweigh the profes- Christian Robert sional aspects. Taking a post-doc job is a big li- [email protected] fe decision with a number of important conside- rations like: Will you be excited about the work Postdoctoral positions are an essential part of you’ll be doing? Who will you be interacting a researcher training that I recommend to all my with? Will you connect with these people? Will PhD students, even though they do not always you have fun? Will this position/location work follow this recommendation! Especially in our for you socially? Will your spouse be happy? Etc. French academic structure where (a) PhD stu- After 4+ years of grad school, you’re finally in dents complete their PhD at 25 or 26, (b) the charge of what you’ll be doing. Hopefully you job market is rather favourable to statisticians, know yourself well enough to decide if a parti- (c) regular university positions are tenured from cular postdoc will be a good fit. Lots of people the start, but (d) involve a heavy teaching load... will have lots of good advice and notions of what Unless family issues are preventing one from makes a good career. But in the end, it’s your job spending one or two years abroad, the expe- to make your career into something you want. rience brought by spending this time in a for- Don’t forget to have fun along the way! eign department with a different academic cul- ture is almost invariably highly positive. And this for many reasons: a high level of freedom Fernando Quintana and Alejandro Jara and time for conducting research, writing pa- [email protected] pers and fattening one’s vita, no administrative [email protected] burden, usually in a prestigious institution with a top quality research group, the opportunity to There is no one-size-fits-all answer here. On start collaborations with more senior researchers, one hand, the answers to these questions lie on most often no teaching, sometimes the opportu- the student’s interests and potentialities. On the nity of learning a new language, the possibility other hand, not all post-docs are the same. It of discovering a new country, etc. Even though seems difficult to derive a general rule for such sabbatical years are available in most academic recommendations, so we limit ourselves to dis- systems, this is clearly the freest of all times in cussing a few relevant elements that can be hel- one’s life and taking the opportunity of a post- pful in this matter. doc can shape one’s academic and non-academic If the student is motivated by a research future! So, unless there are no postdoc positions and/or academic career, then the recommenda- available anywhere appealing (!), or those offe- tion is to definitively do a post-doc. Indeed, this red are in a topic that sounds too far from one’s would be a great opportunity to boost the re- PhD research, or under conditions that are too search line started in the dissertation (or to try constraining, there are few reasons to miss the new or complementary ones), publish all the re- benefits and the fun of doing a postdoc.

Content 14 www.bayesian.org ISBA Bulletin, 18(2), June 2011 STUDENTS’ CORNER

To the question ... well, it depends on the in- dividual of course. Professional maturity is the Stephen Walker key; this operates on a different time scale for [email protected] each of us while having multiple aspects that weight differently for each of us. This – to me A PhD now is not a commitment to an aca- – is the main point for consideration in looking demic career, but a post-doc surely is. Giving at postdoc positions versus “real jobs” on gra- recommendations to someone about any care- duation: the role of the PhD advisor here is to er choice is fraught with difficulties and issues. help students to honestly understand how “rea- My general attitude is not to suggest any career dy” they are to take on the challenges inherent path to anyone (save my own children and pa- in each of the opportunities they may have. So- rents), but to give full support to someone who me students are obviously “ready” for a faculty is keen and knows what it is they want to do. or comparable non-academic position on gra- But to answer the question. I would recommend duation; students having had broader exposure my student do a post-doc if on some occasion I to multiple research areas, interdisciplinary as was asked “Look what I’ve done – what do you well as “core” statistics projects, and a range of think?” And I see something I wished I had done teaching experiences can and do move seamles- myself or thought of myself. [Or some equivalent sly into academic research/teaching positions, or event to this]. And for the second part. When is positions with a comparable (broad!) spectrum a post-doc not appropriate? When the passion is of demands on intellect, energy and professional not there. When the ability is not there. But the maturity in industry, government labs or else- person who would know this the best is the stu- where. Other students may simply be less broad- dent themself. ly experienced and might more clearly benefit from what is typically a relatively relaxed and less demanding postdoc research position; post- Mike West docs can provide an immensely positive personal “time to breath” as well as the opportunities for [email protected] professional and research growth. In other cases, Until the beginning of the 1990s, postdocs we- a student may be quite ready for a “real job” but re almost unheard of in mainstream statistics attracted to the opportunities a postdoc offers in (though more common in biostatistics). As sta- new research directions, to broaden (or, in some tistics grew “out” into this modern interdisci- cases, move away from!) thesis research areas, plinary age, with more aggressive involvement as well as the opportunities to work in a diffe- of stochastic modelling and model-based statisti- rent research group/environment and with new cal reasoning in larger, multi-faceted and longer senior mentors. Postdocs in statistics – whether term research programs, professional develop- “basic” research and in large-scale interdiscipli- ment evolved with substantial expansion of post- nary environments – serve an increasingly active doc opportunities. Nowadays, graduating PhDs and substantial role here, and, for me, an immen- have opportunities to continue their research sely valuable one for the profession. Thankfully, education, statistical and scientific exposure and the historical academic emphasis on moving new broadening, and their early entree into research PhDs as fast as possible into prime academic po- leadership and the profession with postdocs in sitions – to the benefit of the PhD advisor and a range of environments: from core statistics in department “placement” record but sometimes academia, to interdisciplinary statistics in acade- not so obviously to the new PhD in the very early mia, government labs and industry, and in natio- years – is less evident nowadays; broad ranges of nal institutes of various flavours. It is a continu- postdoc opportunities are positive, desirable and al delight for me – and should be for all of us – increasingly important components of the pro- that our professional world is so robust and dy- fessional infrastructure of statistical science, and namic, and that the “sellers marker” for statisti- I hope we have by now grown out of the histori- cians – especially Bayesians with strong applied cal view of postdocs as “second class” next-steps and computational interests – is arguably unique for new researchers. in the range of exciting and rewarding early career opportunities; and I see nothing but continua- tion for decades to come.

Content 15 www.bayesian.org ISBA Bulletin, 18(2), June 2011 NEWS FROM THE WORLD

Dissertation Abstracts rithms, there is a rich literature on shrinkage and variable selection methods for high dimensional BAYESIAN SPARSE LEARNING regression and classification with vector-valued FOR HIGH DIMENSIONAL DATA parameters, such as lasso and the relevance vector machine. Comparing with the Bayesian sto- by Minghui Shi chastic search algorithms, these methods does [email protected] not account for model uncertainty but are more Department of Statistical Science, computationally efficient. In the second part of Duke University this thesis, we generalize this type of ideas to ma- PhD Supervisor: David Dunson trix valued parameters and focus on developing efficient variable selection method for multiva- In this thesis, we develop some Bayesian spar- riate regression. We propose a Bayesian shrinka- se learning methods for high dimensional data ge model (BSM) and an efficient algorithm for analysis. There are two important topics that are learning the associated parameters. related to the idea of sparse learning – variable In the third part of this thesis, we focus on the selection and factor analysis. We start with Baye- topic of factor analysis which has been widely sian variable selection problem in regression mo- used in unsupervised learnings. One central pro- dels. One challenge in Bayesian variable selec- blem in factor analysis is related to the determi- tion is to search the huge model space adequa- nation of the number of latent factors. We pro- tely, while identifying high posterior probabili- pose some Bayesian model selection criteria for ty regions. In the past decades, the main focus selecting the number of latent factors based on has been on the use of Markov chain Monte Car- a graphical factor model. As it is illustrated in lo (MCMC) algorithms for these purposes. In the Chapter 4, our proposed method achieves good first part of this thesis, instead of using MCMC, performance in correctly selecting the number of we propose a new computational approach ba- factors in several different settings. As for appli- sed on sequential Monte Carlo (SMC), which we cation, we implement the graphical factor model refer to as particle stochastic search (PSS). We il- for several different purposes, such as covarian- lustrate PSS through applications to linear regres- ce matrix estimation, latent factor regression and sion and probit models. classification.▲ Besides the Bayesian stochastic search algo-

NEWS FROM THE WORLD

CALLFOR ANNOUNCEMENTS to, Japan. See the June 2010 issue of the ISBA Bulletin for the announcement and more infor- Sebastien Haneuse mation, as well as the conference website http: //www2.e.u-tokyo.ac.jp/∼isba2012/. [email protected]

I would like to encourage those who have any Meetings and conferences announcements or would like to draw attention to an up-coming conference, to get in touch with me and I would be happy to place them here. III Latin American Meeting on Bayesian Stati- stics & XXXVIII National Meeting of Statistics, Pucon,´ Chile. 23-27th October, 2011. Announcements The III COBAL is an international event whose purpose is to allow discussion and interacti- 2012 ISBA World Meeting on among researchers, academics and students, Planning has already begun for the 11th ISBA with the aim to further strengthen scientiﬁc ex- World Meeting, to be held in June 2012 in Kyo- change within the Bayesian community in La-

Content 16 www.bayesian.org ISBA Bulletin, 18(2), June 2011 NEWS FROM THE WORLD tinamerica. This Congress has been performed lace and Maiden’s Tower. twice. The first time was in 2002 in Ubatuba, Bra- Additional information can be found at www. zil, and the second time in 2005 in Los Cabos, Me- worldcong2012.org. xico. This year 2011, the University of Santiago de Chile is leading the organization of the event in collaboration with other sponsoring universi- Short courses and workshops ties. The joint event will take place specifically in Pucon´ Headquarters at the University of La Fron- tera, University Partner in the organization, whe- Bayesian Inference in Stochastic Processes re the series of talks and conferences are schedu- Workshop, Getafe (Madrid), Spain. 1-3rd Sep- led. Spanish and English are the official langua- tember, 2011. ges. In this workshop, we will bring together ex- Additional information can be found at http: perts in the field to review, discuss and explore //cobal2011.usach.cl/. directions of development of Bayesian Inference 8th World Congress in Probability and Stati- in Stochastic Processes and in the use of Stocha- stics, Istanbul, Turkey. 9-14th July, 2012. stic Processes for Bayesian Inference. Theoretical and applied contributions to any area of Bayesi- Jointly organized by the Bernoulli Society and an inference for stochastic process will be welco- the Institute of Mathematical Statistics and scheme. The workshop will be held in an informal en- duled every four years, this meeting is a ma- vironment to encourage discussion and promo- jor worldwide event for statistics and probabili- te further research in these fields. The workshop ty, covering all its branches, including theoreti- will be located in the Universidad Carlos III de cal, methodological, applied and computational Madrid, in the Getafe campus, less than a 20 mi- statistics and probability, and stochastic proces- nute train journey from the centre of Madrid. ses. It features the latest scientific developments Additional information can be found at http: in these fields. //www.est.uc3m.es/bisp7/. The program will cover a wide range of topics in statistics and probability, presenting recent developments and the state of the art in a variety Bayes-250 Workshop, Edinburgh, Scotland. 5- of modern research topics, with in-depth sessi- 7th September, 2011. ons on applications of these disciplines to other The Schools of Mathematics and of Informatics sciences, industrial innovation and society. It will at The University of Edinburgh are holding a re- feature several special plenary lectures presented search workshop to mark the 250th anniversary by leading specialists. In addition, there will be of the death of Thomas Bayes, a former student of many invited sessions highlighting topics of cur- the University of Edinburgh. The workshop will rent research interests, as well as a large number run from early afternoon on Monday 5th to late of contributed sessions and posters. morning on Wednesday 7th September. The ge- The venue of the meeting is Grand Cevahir neral theme of the workshop is that of what has Hotel & Convention Center located in Istanbul come to be known as Bayesian statistics. Places which is a vibrant, multi-cultural and cosmopoli- at the workshop are limited, but there are still a tan city bridging Europe and Asia. Istanbul has a number of places available. unique cultural conglomeration of east and west, For further information see the workshop offering many cultural and touristic attractions, web site http://conferences.inf.ed.ac.uk/ such as Hagia Sophia, Sultanahmet, Topkapo˜ Pa- bayes250.▲

Content 17 www.bayesian.org ISBA Bulletin, 18(2), June 2011

Executive Committee Board Members:

President: Michael I. Jordan 2011–2013: Ming-Hui Chen, Chris Hol- Past President: Peter Muller¨ mes, Antonietta Mira, Judith Rousseau President Elect: Fabrizio Ruggeri 2010–2012: Siddhartha Chib, Arnaud Dou- cet, Peter Hoff, Raquel Prado Treasurer: Hedibert Lopes 2009–2011: David Dunson, David van Dyk, Executive Secretary: Merlise Clyde Katja Ickstadt, Brunero Liseo

Program Council

Chair: Igor Pruenster Vice Chair: Vanja Dukic Past Chair: Alex Schmidt

EDITORIAL BOARD

Editor

Manuel Mendoza [email protected]

Associate Editors

Interviews Students’ Corner Donatello Telesca Luke Bornn [email protected] www.stat.ubc.ca/∼l.bornn/ [email protected] Annotated Bibliography Beatrix Jones News from the World www.massey.ac.nz/∼mbjones/ Sebastien Haneuse [email protected] http://www.centerforhealthstudies.org/ ctrstaff/haneuse.html Software Highlight [email protected] Alex Lewin www.bgx.org.uk/alex/ [email protected]

Content 18 www.bayesian.org