The Nominations Committee was appointed by the Board under the direction of President John Geweke. The members of the Committee were Susie Bayarri, Chair (University of Valencia, Spain), Kathryn Chaloner (University of Minnesota, U.S.A.), Daniel Peña (Carlos III University, Spain), Raquel Prado (Simon Bolivar University, Venezuela), Jim Press (University of California, U.S.A.) and Fabrizio Ruggeri (CNR-IAMI, Italy).
We began the first contacts at the beginning of September and we established the rules to follow in the nominations and voting. In the next elections, we shall be choosing the President Elect and four members of the Board. The Nominations Committee is thus presenting two candidates for President Elect and eight candidates for members of the Board, so as to meet the requirement in ISBA Constitution. The candidates (in alphabetical order) are:
For President Elect: Alicia Carriquiry (USA) Luis Pericchi (Venezuela)
For Members of the Board: Deborah Ashby (UK) Dani Gamerman (Brasil) Eduardo Gutierrez-Pena(Mexico) Irwin Guttman (USA) David Heckerman (USA) Elias Moreno (Spain) Dalene Stangl (USA) Mark Steel (UK)
You will find a short description of the candidates for President Elect and for board members below. Each candidate was asked to provide: affiliation, current status, web page, areas of interest (up to 4), most important journals (or books) in which s/he has published (up to 6) and previous services to ISBA. Candidates for President Elect could add two areas of interest and another two journals/books, as well as Honors received and positions held in other societies.
From here I want to express my most sincere thanks to all the members
of the Nominations Committee for their inputs, comments, and
enthusiastic response. They have done a most thoughtful and responsible
job, and made my Chairing a smooth and easy task. I would also like to
thank the candidates for their fast and positive responses, and for being
willing to devote a most needed portion of their time and efforts to
contribute to the growth of ISBA. I would like to invite you all to
participate in the next elections.
~dalene). As an assistant professor she has co-edited two books on Bayesian methods: Bayesian Biostatistics, 1996, Marcel Dekker and Meta-Analysis in Medicine and Health Policy, to appear late 1999 or early 2000, Marcel Dekker. Her professional interests are hierarchical survival models, decision analysis, and the reform of statistical-education and statistical-practice. In addition to the two books mentioned, recent statistical publications have appeared in Statistics in Medicine, Sankhya, and Lifetime Data Analysis as well as substantive research journals in medicine and health policy. She served as the '98 program chair for the ASA Section on Bayesian Statistics.
~msteel/. He has served as ISBA Vice Program Chair in 1997, Program Chair in 1998, and is currently Past Program Chair.
The ISBA 2000 meeting in Crete, co-sponsored by EUROSTAT and the Association of Balkan Statisticians, looks set to start off the new Bayesian millennium with a bang. Our hard-working Programme Committee has been inundated with high-quality proposals for sessions and speakers, and an exciting conference of the highest scientific interest and quality seems assured. The programme will feature a wide range of up-to-the-minute theoretical and methodological advances over the whole field of Bayesian Statistics. The power of Bayesian thinking and analysis to further our understanding of the world will be demonstrated across a broad swathe of applied problem areas. And there will be a special emphasis on highlighting opportunities for interdisciplinary interaction. A number of sessions will be devoted to the systematic exploration of the potential of the Bayesian approach for solving important practical problems of government, official statistics and public policy. The Bayesian approach has a substantial contribution to make towards meeting the challenges faced by Official Statistics, as it strives to balance calls for more and better data against the costs and burdens of providing these services. To this end, ISBA has decided to make Official Statistics a leading theme of its Sixth World Meeting, with the following aims:
a) To bring together Bayesian scientists and official statisticians, to exchange information and views on the state of the Bayesian art.
b) To identify problem areas in official statistics, broadly interpreted, where the application of Bayesian tools and methods appears most promising.
c) To lay the foundations for collaborative research to address issues of concern.
To address these aims, as well as similar issues in other areas, the meeting will tackle fundamental implementational issues such as the role and validity of subjective judgement in the construction of models and priors, the feasibility of the elicitation process, and computational practicalities. It will take stock of a wide range of methodological tools, such as MCMC computation, model-averaging, influence diagrams and belief nets, and hierarchical modeling. It will present and study late-breaking original new developments in Bayesian theory and methodology. And it will explore a wide range of actual and possible applications, especially where relevant to the theme topic of official statistics -- whether generic, or in particular domains such as health education and environment. We look forward to welcoming you to Hersonissos next summer.
Program planning for ISBA 2000 approaches completion. Following an amazing, world-wide response to the call for proposals earlier this year, the Scientific Committee has now finalised the formal scientific program.
ISBA 2000 will feature:
Over 40 technical sessions of oral presentations, featuring more than 120 individual talks, and 3 evening poster sessions featuring a similar number of individual poster presentations.
A range of "Theme" related sessions covering applied Bayesian work in official statistics and policy arenas, and challenges and opportunities in these areas. In addition to official, policy and governmental statistics, a further subset of sessions is devoted to Bayesian methods in public policy more broadly, including public health, environment and agriculture.
A wide range of "General" sessions of talks on Bayesian theory, methodology and application in many diverse areas - highlighting current research frontiers, and, consistent with one of ISBA's primary institutional goals, showcasing important interdisciplinary applications of Bayesian methods. There will be something for everyone, and much more than enough for most.
An incredible range of topics
ISBA Newsletter, September 1999 ISBA 2000
In the very near future, full details of scheduled oral and poster sessions
will be listed on the ISBA 2000 web site,
This can also be accessed through the ISBA web site www.bayesian.org and provides all conference information, including registration and travel, in addition to the scientific program.
We invite you to browse the web site, register for ISBA 2000, and encourage
your colleagues, students and friends to do so too. We anticipate an unusually
active, informative and productive meeting, and look forward to seeing fellow
ISBA members and many other Bayesian colleagues in Crete next May.
228 Location and Venue
The ISBA 2000 Conference will be held at Hersonissos, a popular summer resort on the north coast of Crete, just 25 km from Heraklion International airport. The Knossos Royal Village and the Royal Mare Village hotels have been selected as Conference Sites. Both of them are luxury class purpose built villages. The Conference Hall belongs to the Knossos Royal Village. It has all the required facilities and it has been the venue of international congresses many time in the past. The two hotels are next to each other. The Royal Mare Village hotel can be considered as more luxurious and modern but delegates may need to walk up to 6 min. to reach the Conference Hall from their rooms.
228 Social events and Local Information
Trips are being organised to Knossos, Heraklion Archaeological museum and Matala beach.
Crete is the largest of the Greek islands and the fifth in size of all the islands of the Mediterranean Sea. It is also the most southern point of Europe. It lies at the Southern Aegean Sea and at the crossroads of three continents Europe, Asia and Africa. Crete covers an area of 8336 sq.km. The length of the island is 260 km, but the shore-length is 1046 km. A high mountain range crosses the island from West to East, formed by three different groups of mountains. To the West the White Mountains (2452 m), in the middle the mountain of Idi (2456 m) and to the East the mountain of Dikti (2148 m). These mountains gifted Crete with fertile plateaus like Lasithi, Omalos and Nidha, caves like Diktaion and Idaion and the famous Gorge of Samaria. The Gorge of Samaria is the longest in Europe, measuring some 18-km and can be covered in about 7 hours on foot. It is well known for its awesome beauty. At some points the passage is just 3 meters wide and at times the steep sides rise to a height of 600 meters. A stream that flows cuts the gorge between the highest peak of the White Mountains and the mountain of Volikas.
A place with a great history from ancient times, Crete offers visitors a priceless wealth of findings of all the civilizations that flourished on the island in its museums and in its archeological sites. Knossos contains the ruins of the largest and most luxurious Minoan palace, built in the middle of a large town. The impressive Minoan Palace of Knossos is famous throughout the world for its association with the myths of the Minotaur, the Labyrinth, Daedalus, Ikarus and of course of Theseus, and the most ancient civilization in Europe.
The Archaeological Museum is one of the most outstanding museums in the world. It contains findings from all over Crete, focused primarily on the prehistoric Minoan civilization that ruled the island for over 1200 years.
Matala was the port of Phaistos during the Minoan period. Ruins of the
ancient city are still visible on the seabed as the ancient city was
sunk in the sea. Due to its exceptional natural beauty, Matala became
the meeting place of the "Flower Children" in 1968. Although their
conference failed to realize, yet they were compensated by the
incomparable beauty of the area, which so much contrasts with the
concept of destruction and war.
ISBA Newsletter, September 1999 INTERVIEWS
Peter Müller is an Associate Professor at the ISDS in Duke University. He has produced innovative work in a number of areas of Bayesian Statistics, in particular in Bayesian non-parametrics and utility and decision problems.
Part of this interview was conducted live with our editor in
attendance before the wedding of our co-interviewer David Rios
Insua. Congratulations to David for finally using his Bayesian
judgements to get married to someone who doesn't have anything to do
1) Why did you decide to become a statistician?
Like many statisticians, probably by coincidence. I went to the US for
graduate studies having taken Math/Phys in Austria, so statistics
seemed like a reasonable choice, if not a conscious choice.
And why Bayesian? There aren't too many Bayesians in Austria
As I came to Purdue to study statistics I met many Bayesians there. Jim Berger was a big influence.
As for Austria, it is not exactly a Bayesian free zone. There was Wolfgang Polasek at the University of Vienna at that time. And there are now Klaus Pötzelberger, Klaus Felsenstein, Frühwirth-Schnatter and colleagues at the Wirtschaftsuniversität Wien.
2) Who are the people who have had the strongest influence on
your statistical career?
Jim Berger. I studied for my doctorate with him at Purdue and he gave
me some good ideas on what's important in statistics. I can identify
very strongly with his views on reconciling Bayesian statistics with
practical analyses. Although he does not work a lot in computational
methods, he was very supportive. It's also because of him
(indirectly, as he had a sabbatical there and I went with him) that I
ended up at Duke. While in the sabbatical year there I took a class
with Mike West which greatly impressed me. Also John Geweke has
taught me a lot. Meeting them helped to push me further down the road
of computational Bayes. Also this was just the right time for
starting to work in computational Bayes, just year 2BG (before Gibbs).
3) You have worked on a number of areas of Bayesian
non-parametrics; Dirichlet processes, wavelets, neural nets etc. Can
you tell us something about this work?
The work on Dirichlet processes (DP) has not been me alone, but many people have contributed. My interests were mainly in applying DP models to parametrize mixture models used to generalize fully parametric models for random effects distributions, measurement error models; non-linear autoregression, etc. Also, the problem of implementing posterior simulation for non-conjugate DP mixtures kept us puzzling for a while.
As for wavelets, having an office only two doors down from Brani Vidakovic, I have to be interested in them. They seem a very natural way of modelling functions that leads to more robust modelling opportunities. It seems a safe guess to assume that wavelet transformation will become as commonly used a tool in statistical inference as Fourier decomposition is now.
I see wavelet based modeling as the exact opposite of another
non-parametric modeling approach which interested me a while ago,
namely neural networks. While wavelets provide parsimonious
parametrization of random functions, neural networks follow the exact
opposite paradigm and throw intentionally massively overparametrized
models at the data. It still intrigues me trying to understand what is
the correct way of approaching such models.
4) What do you think is your most important contribution to
I wouldn't like to pick out any single thing. I think I've made many
small contributions to computational problems, though nothing
Some of the more fun things to work on were an algorithm for posterior
simulation for non-conjugate Dirichlet process mixtures (MacEachern
and Müller 1998),
semi-parametric models for longitudinal data (Müller and
Rosner 1997), Bayes in case-control
ISBA Newsletter, September 1999 INTERVIEWS
simulation based approaches to decision problems (Bielza, Müller and Rios Insua 1999).
5) What are the major advances in statistics you've seen in your
Well, it has not been that long! In Bayesian computation, Gibbs sampling, Tierney's MCMC paper.
6) And if you had a crystal ball to look into the future?
I can answer only from my very subjective perspective -
going beyond traditional parameter estimation to solving decision
problems, moving towards more meaningful problems, clinical trials etc.
7) What about your own future research plans?
I'd like to work on some interesting decision problems, extending some of the methodology which we have developed over the last 10 years for parameter estimation. In many computational approaches, there are an infinite number of possible extensions, but it is often very hard to focus without a meaningful practical application.
8) Do you have any advice for teachers of Bayesian statistics?
I don't have any real tips. My main advice would be not to have excessive expectations and to keep things in perspective. I do find computers useful in teaching though. I use the web a lot and email to address concerns in large classes.
9) What do you enjoy most about your work?
It's fun to work with people from other fields. I think the level and diversity of interdisciplinary collaborations in statistics in general, and Bayesian statistics in particular, is quite exceptional compared to many other fields. On the less academic side, being able to visit collaborators or go to conferences in nice places is often a treat. Bayesian statisticians are a friendly bunch.
Working on revisions for manuscripts when you don't at all agree with referees, but have to somehow try to read sense into the comments. And on the other side, I find it very painful to referee bad papers.
10) What is your favourite statistics book?
Some random answers; Jim's book, Chris Robert's Bayesian Choice and Brani's wavelet book; Thisted's Computational Statistics book.
11) What is your favourite Bayesian statistics joke?
I don't really know any Bayesian statistics joke. Here is the closest
I can think of.
You want to find the solution to 1 + 1. Ask an engineer and he'll
tell you 1 + 1 = 2, ask a statistician and he'll say 2
0.1, ask a
mathematician and he'll go away and come back in a few days and tell
you that the solution exists. Finally, if you ask an accountant,
he'll close the door and ask you how much you want it to be.
12) As a member of ISBA, what are your views about the Society and what, if any changes would you like to see?
I know some people are sceptical about a society based on methodological focus but I see nothing wrong with that, especially as I see that the next conference has a very specific theme. I identify strongly with the aims of ISBA such as encouraging collaboration, support for young researchers etc. The focus of ISBA I like most is it's international orientation. I'd like to see this continue, drawing in more people from diverse backgrounds and different countries.
References given in Peter's talk are ...
which converges when | x1| < 1,...,| xn| < 1. Note that (a, n) : = for any positive a and integer n.
Lauricella obtained interesting integral representations for his functions.
As far as FD(n) is concerned, he found that
(1) FD(n)(a, b1,..., bn;c;x1,..., xn)=fn(u1,..., un)(1-u1x1-...-unxn)-adu1...dun
holds true for Re(b1),...,Re(bn),Re(c - b1 -...- bn) > 0 with
Moreover, Lauricella proved that
(2) FD(n)(a, b1,..., bn;c;x1,..., xn) = ua - 1(1-u)c - a - 1(1-ux1)-b1...(1-uxn)-bndu
is valid when Re(a) and Re(c - a) are strictly positive. Thus
(3a) fn(u1,..., un)(1-u1x1-...-unxn)-adu1...dun
For example, some of the Lauricella classical results allow us to obtain the expression of the Stieltjes transform (of order c) of in a much more direct and simple way than that originally exhibited in Cifarelli and Regazzini (1990). Indeed, after indicating that transform by , we have
What is the bare minimum that someone interested in applying Bayesian
methods must know, just to get started? This is the first question we
have asked ourselves in the context of both a half-semester class for
first-year graduate students and a one-day short-course for
practicing statisticians. A second question follows closely on the
first: What difficulties will students have in trying to learn these
essentials in such compressed time-frames?
Our strategy has been to try to emphasize (i) the nature of Bayesian
data analysis, and (ii) the specific technique of hierarchical
modeling. Along the way we talk about (iii) the choice of prior
distributions, (iv) computation (mainly, posterior simulation), and
here and there (v) a little philosophy.
It is of course important to keep in mind the great effectiveness in
Statistics (as well as in other fields) of teaching via examples: they
make concepts concrete, and force the instructor to follow through
with claims of how the Bayesian approach is supposed to provide
insight. In this introductory context we keep things as simple as
possible: all computations are done in S-PLUS.
We begin with a one-sample Binomial example, partly because we have
found that students rarely question the use of a uniform prior for the
Binomial parameter--we question it ourselves, but not at the outset.
In the context of the Binomial, we talk about data-domination in large
samples, and the way that a Normal distribution then provides a good
approximation to the posterior, even when it is based on Maximum
Likelihood. We consider this crucial pedagogically: it is helpful to
students to make a connection with something familiar, and they must
also understand the similarities and distinctions between Bayesian and
We next talk about the use of marginalization (integration) to
eliminate nuisance parameters, and we illustrate with an example in
which we compare two Binomials. Only after presenting results and
their interpretation do we ask how this integration may be carried
out. We describe posterior simulation in this very simple case (where
we can generate the joint posterior--a product of independent
Betas--directly). We then review some of the appealing aspects of
Bayesian inference, including its intuitiveness, which we illustrate
with the example of interpreting a thermometer reading, taken from the
famous 1963 paper by Edwards, Lindman, and Savage. We also review in a
very cursory manner some basic foundational notions, including the
meaning of probability and the indeterminacy of confidence, and we
point out that, in applications, Bayesian inference is often used in
conjunction with frequentist (often exploratory) methods.
It is our view that what we've described so far constitutes the core
of the material, yet we can move through it quickly: in just two or
perhaps three classes in the case of the half-semester course, or a
couple of hours when we're whipping through the day-long version. On
the other hand, everything else should reinforce these basic ideas,
and we say so as we progress through the remainder.
A couple more comments may help. Our attitude about MCMC is that in
principle students really shouldn't have to know much about it, but in
our current reality they must have a basic understanding. So we remind
them of the basics of Markov chains (in the discrete case) and go
through the Metropolis algorithm and Gibbs sampling. We use
examples involving only a few parameters, and we derive the Metropolis
algorithm constructively, thereby answering the natural question, How
did anyone think it up? Then we go on to Hierarchical modeling, where
we present the simplest Normal-Normal model in great detail,
discussing both empirical Bayes and fully Bayes approaches, together
with interpretation of a simple data set. We briefly indicate that the
power of all this stuff comes from our ability to attack complicated
problems, but in fact we're done: the course is already over. Those
who wish to pursue more interesting examples will, we hope, be
reasonably well-equipped to do so on their own.
We are trying to finish a book that supports this minimalist
presentation. It is called A Short Course on Bayesian Statistics
and should be available from Springer next August.
ISBA Newsletter, September 1999 BIBLIOGRAPHY
Statistical reliability theory is the study of the failure of systems through probability. By defining 'system' in a broad sense, reliability can therefore include such diverse topics as fault tree analysis, failure lifetime modeling, product development and testing, aspects of biometrics, risk analysis and insurance. Here, some recent papers in several different areas of reliability that use Bayesian methods are given.
Papers on reliability appear in all the major statistical journals, mainly in the applications section (i.e. more likely Applied Statistics than JRSS B). The IEEE Transactions on Reliability is another good source, as is the IEEE Transactions on Software Engineering for papers in software reliability and testing.
Although not particularly Bayesian, the classic methodological text in reliability is
[0.7mm] R. E. BARLOW AND F. PROSCHAN(1981). Statistical Theory of Reliability and Life Testing , 2nd edition, Silver Spring.
There are two volumes, the first dealing with models and the second with inference. Perhaps a better place to start is
[0.7mm] N. D. SINGPURWALLA (1988). Foundational issues in reliability and risk analysis, SIAM Review, vol. 30, no. 2, pp. 264-282,
which lays out the foundational issues in reliability from a Bayesian perspective, heavily motivated by the work of de Finetti. Like all areas of Bayesian application, Markov Chain Monte Carlo has considerably widened the class of models that can be practically used, and has also allowed the issue of prior choice and specification to be addressed more satisfactorily. This is one strength of Bayesian methods in reliability, particularly engineering applications where there is usually a large body of expert opinion. A good example of prior elicitation with application to reliability testing is:
[0.7mm] G. A. WHITMORE, K. D. S. YOUNG AND A. C. KIMBER(1994). Two-stage reliability tests with technological evolution: a Bayesian analysis. Applied Statistics, vol. 43, no. 2, pp. 295-307.
Closely related is the area of product development. The following paper demonstrates that Bayesian methods are easy and natural to use in this area also.
[0.7mm] T. A. MAZZUCHI AND R. SOYER(1993). A Bayes method for assessing product reliability during development testing. IEEE Transactions on Reliability, vol. 42, no. 3, pp. 503-510.
The proportional hazards model of Cox is probably the most widely used approach to lifetime modeling in the presence of covariates. A good reference using this model via the Bayesian approach, is:
[0.7mm] C. T. VOLINSKY, D. MADIGAN, A. E. RAFTERY, AND R. A. KRONMAL(1997). Bayesian model averaging in proportional hazards models: assessing the risk of a stroke. Applied Statistics, vol. 46, no. 4, pp. 433-448.
There is a good description of how to apply the proportional hazards model in a Bayesian setting here.
Another area of reliability that is provoking a lot of interest is software reliability and testing. The following book sets out the Bayesian approach to software engineering in general, as well as devoting considerable space to software reliability and testing.
[0.7mm] N. D. SINGPURWALLA AND S. P. WILSON(1999). Statistical Methods in Software Engineering: Reliability and Risk. Springer-Verlag, New York.
Some more articles with specific applications are given below.
Development and analysis of both
attribute- and variable-data reliability growth models are covered in the following paper.
[0.7mm] A. ERKANLI, T.A. MAZZUCHI AND R. SOYER(1998). Bayesian computations for a class of reliability growth models. Technometrics, 40, pp. 14-23.
This paper begins with an overview of a Bayesian attribute-data reliability growth model and illustrates how this model can be extended to cover the variable-data growth
ISBA Newsletter, September 1999 BIBLIOGRAPHY
In the article below, a Bayesian approach is developed for determining an
optimal age replacement policy with minimal repair.
S.-H. SHEU, R.H. YEH, Y.-B. LIN, AND M.-G. JUANG(1999). A Bayesian perspective on age replacement with minimal repair. Reliability Engineering and System Safety, 65, pp. 55-64
By incorporating minimal repair, planned replacement, and unplanned replacement, the mathematical formulas of the expected cost per unit time are obtained, and it shown that there exists a unique and finite optimal age for replacement. Using the Weibull distribution with uncertain parameters for failure time, the above article develops a Bayesian approach to formally express and update the uncertain parameters for determining an optimal age replacement policy.
Two statistical approaches to on-line prediction of
cutting tool life are presented and discussed in the following article.
[0.7mm] H. WIKLUND(1998). Bayesian and regression approaches to on-line prediction of residual tool life. Quality and Reliability Engineering International, 14 (5), pp. 303-309.
A Bayesian approach utilizing in-process information about cutting tool state, and a second approach based on the cutting forces are presented in the above article.
The following article uses Bayesian approach to develop significance tests for testing for a reduction of the
Modulated Power Law process(MPLP) to simpler models, namely the Gamma Renewal and the Power Law
processes, which are special cases of the MPLP model.
[0.7mm] R. CALABRIA, G. PULCINI(1999). On testing for repair effect and time trend in repairable mechanical units. Communications in Statistics - Theory and Methods, 28 (2), pp. 367-387.
Furthermore, significance tests to compare the effect of the repair actions on the future reliability or the time trend in two independent MPLP samples are proposed and studied in the above article.
The Galileo spacecraft deployed a probe, during 1995, to investigate the
atmosphere of Jupiter and it was powered by Li/SO2 batteries. The
fundamental problem for the decision-makers during the mission was the
uncertainty in knowing whether the batteries had sufficient capacity
left to perform the planned mission. Accounting for all
identified uncertainties, the following article employed a Bayesian Weibull analysis using a Monte
Carlo solution technique, and determined the confidence that the battery set
on-board the Galileo probe would perform adequately.
[0.7mm] M.V. FRANK, K. SILKE (1998). Galileo-Probe battery-lifetime estimation. 1998 Proceedings of the Annual Reliability and Maintainability Symposium, pp. 76-81.
In the following report, a general class of Accelerated life tests(ALT) models are proposed, which are motivated by actual failure process of units from a limited failure population with a positive probability of not failing during the technological lifetime.
[0.7mm] D. SINHA, K. PATRA, AND D. DEY(1999). New Bayesian Approaches for Accelerated Life Test Data. Technical Report, Department of Statistics, University of Connecticut. Correspondence: firstname.lastname@example.org.
The following paper deals with computational techniques to estimate the parameters
and the reliability function of complex life distributions using complete and Type-II
[0.7mm] D. DEY AND T. LEE(1992). Bayes Computation for Life Testing and Reliability Estimation. IEEE Transactions on Reliability, Vol. 41, No. 4, pp621-626.
Thanks to Simon Wilson, Maurizio Guida, Gianpaolo Pulcini and Dipak Dey for
their valuable help with the collection of the references.
ISBA Newsletter, September 1999 STUDENT'S CORNER
We present in this issue four abstracts. Jennifer Hills is a Ph.D candidate under the supervision of Dr.Donald Rubin in Harvard. Jennifer has worked on Bayesian modelling applied to estimation of propensity scores for causal inference. Sarah Michalak is also doing her Ph.D. in Harvard under Dr. Carl Morris. She deals with the problem of ascertaining whether the outcome of patients in different groups are comparable with respect to mortality. Our third contributor from Harvard, Riouxi Tang, is a Ph.D. candidate under Dr. Carl Morris. For the work in the abstract she has worked with Dr. David Van Dyk on improved E-M type algorithms, in terms of faster convergence, in the context of dynamic linear systems. We have our last contribution from Thomas Nichols of Carnegie Mellon University, Pittsburgh. He worked on his thesis with Dr. Christopher Genovese on functional magnetic resonance imaging of the human brain. We are grateful to Sarah Michalak for collecting the abstracts from Harvard and to Dr. Genovese for sending in Nichols' abstract while he was out of town.
The development of Bayesian statistics in Chile started only a few years ago. We all agree that Pilar Iglesias has been the spiritual leader of the group. After working with Carlos Pereira, she received the Doctoral degree from the U. of São Paulo (USP), Brazil. She later became a faculty member in the Department of Statistics of the P. U. Católica de Chile (PUC) in 1994.
From the very day of her arrival at PUC, Pilar started making contact with faculty members at other universities. Consequently, the U. of La Serena, Chile hosted the First Bayesian Workshop, in January of 1996. It was a short event, with a very reduced number of participants, but being the first entirely Bayesian meeting in Chile, it had a special relevance. The workshop came as the result of a joint effort of Pilar, Victor Salinas from the U. de Santiago de Chile (USACH), another former student of C. Pereira in USP, and some local faculty members. Certainly, the whole project was strongly motivated by the Brazilian tradition.
Later, in May of 1996, the U. of Valparaíso (UV) hosted the First Workshop on Models with Errors in Variables and Bayesian Inference. Again, Pilar was at the heart of the organization, sharing responsibilities with Manuel Galea of the UV, and Reinaldo Arellano (PUC), both former students of Heleno Bolfarine in USP. This was also the first time that the Journal of the Chilean Statistical Society (SOCHE) edited a special issue, completely devoted to the papers presented in this meeting.
The Second Workshop was hosted by the U. of Antofagasta (UA) in January of 1997. The local organizers were Héctor Varela, Guillermo Mondaca (former students of Vicente Quesada at U. Complutense de Madrid), and Juan Duarte (former student of Domingo Morales, also at U. Complutense de Madrid), all of them faculty members of the UA.
The Third Workshop was held at the U. Austral de Chile in Valdivia (UAV) in January of 1998, where it was decided that the future versions would take place biannually. The local host was Eliana Scheihing from UAV (former student of Michel Mouchart in the U. of Louvain). A future issue of the SOCHE journal will concentrate on papers presented in Valdivia, this time including discussion. The Third Workshop had the largest international attendance so far, including Manuel Mendoza from ITAM in Mexico, Michel Mouchart and Heleno Bolfarine.
The Chilean chapter of ISBA was established in 1997. We currently have about 25 members from nearly all the universities in Chile, including some local graduate students and some current and former doctoral students in USP. The reader is kindly invited to guess who has been the chair of our organization right from its birth.
The main topics of research conducted by Chilean Bayesians are: de Finetti-type Theorems (Pilar Iglesias); Bayesian modeling with errors in variables and with elliptical errors (Reinaldo Arellano, Pilar Iglesias, PUC; Manuel Galea, UV); nonparametric Bayesian methods (Héctor Varela, Juan Duarte, Guillermo Mondaca, UA; Eliana Scheihing, UAV; Victor Salinas, USACH; Fernando Quintana, PUC); Bayesian modeling (Arturo Mora, María Elena Valenzuela, U. de Concepción; Ernesto San Martín, currently in U. Louvain). We also have to mention Alicia Carriquiry, from Iowa State U., Heleno Bolfarine, and Guido del Pino (PUC) who have repeatedly given us their support and encouragement in our research activities.
Bayesian statistics in Chile is at its early stages, but starting to grow
and develop. Indeed, Pilar and Reinaldo have advised a number of Master
theses at PUC, and some Ph. D. theses at USP. We expect to keep climbing the
ladder in the coming years.
ISBA Newsletter, September 1999 APPLICATIONS
It is perhaps unfortunate that Bayes died before the advent of recorded sound in the mid-nineteenth century. If he had been alive then it might now be possible to enhance a recording of his voice using Bayesian inference, as it has been with early archive material of Queen Victoria, Florence Nightingale and other great figures of history.
This article is concerned with the computer restoration of degraded sound recordings stored on tape, 78rpm disks or even the very early wax cylinder recording media. The Signal Processing Group at Cambridge University has carried out pioneering work in this area over the last fifteen years and details of some of our recent research can be found in the recent Springer book `Digital Audio Restoration' by Simon Godsill and Peter Rayner. In addition, a wide range of processed sound examples and further information can be found at the website www-sigproc.eng.cam.ac.uk/~ sjg/springer.
Research from the Cambridge lab has led to the foundation of a world-leading company, CEDAR Audio Ltd., specialising in restoration of sound recordings for record companies, and there is on-going interaction between academic research and the development of improved commercial algorithms. Recent high profile projects worked on by CEDAR include the remastering of the Star Wars movie sound track.
Audio signal time series are obtained directly from the analogue sound source by careful analogue-to-digital conversion, usually performed at the CD sampling rate of 44.1 kHz and 16-bit resolution to ensure an accurate transcription of wide bandwidth audio signals (high quality music extends in bandwidth beyond 20kHz - above the range of human hearing!). The datasets involved are thus huge, typically requiring the processing of many millions of data points within a single recording. In recent years, however, with the rapid improvements in cheap computing power, it has been possible to incorporate many aspects of Bayesian computational methodology into audio processing algorithms, and it is various aspects of this work within the gramophone field which I will now describe.
There is a wide range of defects which can occur in gramophone recordings, especially when the medium, whether cylinder, disk or tape, has been subjected to wear and tear or poor storage over a period of years. The good news from a Bayesian perspective is that there is usually an abundance of subjective or objective prior information about the mechanisms which degrade a recording.
Take, for example, the characteristic click and crackle noise associated with early 78rpm recordings. We know that these defects are caused by `bumps' and scratches on the surface of the groove walls, some present even when the disk was brand new, and others appearing with time as part of the general ageing process. It is thus anticipated that the corrupting noise will be approximately additive to the musical signal encoded in the groove wall, and will be intermittent or `impulsive' in nature over the audio time series. It is also expected that this noise will act over a wide range of amplitude scales, corresponding to a range of physical defects from microscopic scratches and surface irregularities up to relatively deep gouges or large dust particles adhering to the surface. Thus the distribution of noise amplitudes should be quite heavy-tailed in order to to model all of these features adequately. In fact, these considerations are all borne out by the observed time series, which are clearly seen to have intermittent additive noise disturbances at a wide range of scales. It is these very short bursts of interference that the ear detects as clicks and crackles in an old recording.
We are now in a position to postulate Bayesian models for the noise
processes. The models we have found most successful include a
two-state switching or
ISBA Newsletter, September 1999 APPLICATIONS
Having specified noise distributions, the full Bayesian picture is completed by a model for the musical signal itself. This is a highly complex issue, as there is such a wide variety of possible types. A very sophisticated approach to modelling of musical signals would model each note played by each instrument in terms of its attack, pitch and harmonic series. We then need to add to this considerations of timbre and all possible performance fluctuations. While such models are now being investigated for the challenging task of automatic transcription of sound recordings (essentially, the process of writing down the musical score of a piece of music directly from the raw audio time series), it usually suffices to use a far simpler signal model for the music which captures the salient features for noise reduction purposes. To this end, autoregressive or autoregressive moving-average models have been found to be useful approximations to reality which offer a suitable trade-off between complexity and utility.
Given the full Bayesian model, including priors for unknown hyperparameters in the noise and signal models, the task is now a large scale optimisation for the reconstructed signal values conditional upon the observed data recording. This is far too complex to perform analytically and so we adopt state-of-the-art Markov chain Monte Carlo (MCMC) sampling methods for simulation of the reconstructed data from its posterior distribution. The result is a more accurate detection and elimination of clicks and crackle than was previously possible.
Removal of click and crackle noise is just one aspect of the restoration and analysis of sound recordings. We mentioned earlier the possibility of automated transcription of sound. Here it is crucial to incorporate a very high level model of the sound that includes as parameters the pitches of individual notes played and all of their specific characteristics. This is an interesting problem of variable dimension, since at any time the number of notes playing is unknown a priori. Our initial exploration of this problem, once again using a Bayesian model implemented with MCMC sampling, has given promising results, even for cases where several instruments are sounding simultaneously, so we can anticipate that Bayesian methods will be indispensable in this important area too.
Another problem where the techniques can be usefully employed is the correction of pitch deviation defects, or `wow' as it is known. These can occur through a number of mechanisms, including unevenly stretched tape or warped gramophone disks. In either case, the effect on playback is an unpleasant time-varying pitch variation in the music. Correction of the problem involves modelling of the pitches present in the music and tracking any changes over time of these pitches. Bayesian priors are incorporated for the regularisation of this problem, and can incorporate specific information such as the frequency of variation or very vague information about the expected smoothness of the variation, depending on the degree of our prior knowledge about the the history of a particular recording. Having identified the pitch variations, the final stage in pitch correction involves performing a digital `stretching' of the time axis in order to invert the effect of the identified pitch variations. Results of this procedure are remarkably successful and can be listened to on the web page listed above.
We have summarised here just a few of the applications of Bayesian methods to sound signals. To find details of other audio applications, see the Springer book referenced above. It is hoped that the discussion has highlighted that this is a Bayesian success story, in that
ISBA Newsletter, September 1999 APPLICATIONS / SOFTWARE / NEWS
The Bayesian Output Analysis Program (BOA) is a set of S-PLUS/R functions that carry out convergence diagnostics, statistical and graphical analysis of Monte Carlo sampling output in a similar fashion as CODA, developed by N. Best, K. Coles and K. Vines. The software can be used as an output processor for BUGS or other programs that produce Markov Chain Monte Carlo (MCMC) output. The main features of BOA include menu-driven interface with a stand-alone library of functions. Also, there is a flexible data management which permits analysis of MCMC output in standard format ASCII text files or as S-PLUS/R matrices. The MCMC sequences may have varying lengths and number of parameters. Additionally, BOA is implemented with faster algorithms, less memory usage and corrects for some breakdowns of CODA with S-PLUS version 5.0. Available functions in BOA produce different summary statistics of the MCMC samples, autocorrelations, cross autocorrelations and the same convergence diagnostics of CODA, with the implementation of the Brooks and Gelman multivariate shrink factor approach. It also allows a display of plots that visualize lag-autocorrelations, density estimators for parameters, univariate and multivariate shrink factors, the Geweke convergence diagnostic and individual trace plots. Most of the development for BOA has been done in S-PLUS 5.0 for Linux, but the software has been successfully tested on S-PLUS 3.4 for Unix, S-PLUS 4.0/4.5 for Microsoft Windows; as well as on R 0.64 for UNIX and Microsoft Windows. BOA and related documentation are free for academic purposes and downloadable at:
This conference is the latest in a severals year series of
international gatherings of researchers in the survey field.
However, there has been no international scientific meeting devoted
to survey nonresponse since the early 1980's, and there is no
single printed volume describing these developments, despite
the field has changed in important ways in the 13 years since then.
The conference aims to stimulate the assembly of documentation of
state of the art practice, in order to produce a volume that describes the
state of the art in social science and statistical theory and practice in
nonresponse rate reduction, nonresponse error measurement, and
postsurvey compensation for nonresponse. The volume will not be an ordinary
proceedings book, but an integrated treatment of the field,
also suitable for university use. See the conference web page at
ISBA Newsletter, September 1999 NEWS FROM THE WORLD
Mathematical Methods in Reliability 2000. July 4-7, 2000,
Bordeaux, France. This conference covers a wide range of topics in
reliability, and a session devoted to Bayesian methods
is also included. The deadline for the submission of abstracts is
November 12, 1999; more info at
Knowledge Discovery and Data Mining 2000.
August 20-23, 2000, Boston, MA, USA. This is the
sixth ACM conference on KDD. In
addition to fundamental research, the organizers solicit
papers fostering cross-fertilization and interdisciplinary
integration, as well as papers that describe significant
experiences and implementation lessons (deadline for abstracts:
February 29, 2000; web page: www.acm.org/sigkdd/kdd2000).
228 Internet Resources
ASC software register. The Association for Survey Computing's Register of software for statistical and social survey analysis is browsable through the ASC's web site (www.asc.org.uk). The register content is listed by package, by main function, by feature and by supplier and is quite rich. The main topics that appear in the by-feature list are survey design, data capture, data management, statistical analysis, presentation, and operating system. The information about the software has been provided by various organizations completing the on-line register questionnaire.
228 Research Opportunities
Post-Doctoral Research Position. An opportunity now exists to join a team of researchers in Trinity College Dublin on a project in the field of medical image processing. The project, which is now in its third year, involves researchers from the Departments of Electronic Engineering, Statistics and Medical Physics. The challenge being addressed is the provision of cutting-edge Computer-Aided Diagnosis techniques in a busy gastroenterology referral centre at St. James's Hospital in Dublin. Our expertise in Bayesian methodologies for segmentation of textured images is being employed for this challenging task. In particular, we are emphasizing the development of data-driven Markov Chain Monte Carlo techniques, and Markov Random Field modelling.
The initial contract will be for one year, and will commence as soon as possible. Remuneration will be in the region of 18,000 Irish pounds (22,900 Euros) per year. The successful candidate will take charge of adapting and steering the research in a manner that best addresses the needs of the end-users in medical endoscopy. In particular, they will be responsible for theory and algorithm developments which make possible the implementation of the algorithms in this environment. It is expected, therefore, that the researcher will have a strong background in an appropriate technical area (Bayesian methods, image segmentation), but also have the practical aptitude to work directly on the implementational challenge.
Applicants should contact Dr Anthony Quinn (email@example.com) at
Dept. of Electronic and Electrical Engineering, Trinity College,
Dublin 2, Ireland, providing a CV, a statement of research
background, and the names of three referees.
228 Awards and Prizes
93 Savage Award. Antonietta Mira, U. of Insubria, Italy, won the 1998 Savage Award Competition for her thesis, ``Ordering, Splicing and Splitting Monte Carlo Markov Chains'', completed at the U. of Minnesota under the direction of Luke Tierney. A honorable mention was given to Jaelong Lee, National Institute of Statistical Science, for his thesis, ``Semiparametric Bayesian
ISBA Newsletter, September 1999 NEWS FROM THE WORLD
1999 ASA Outstanding Statistical Application Award.
During the 1999 Joint Statistical Meetings in
Baltimore, the award was conferred on Mike West, Raquel Prado, and
Andrew Krystal for their paper on latent structure in
electroencephalographic traces of depressed patients (JASA, vol. 94).
This is one of a series of articles on statistical time series methods,
motivated by the need to aid clinicians in refining and improving
therapies, as well as in contributing to the understanding of
the underlying neurophysiology.
93 Mitchell Prize 1999.
The 1999 Mitchell Prize has been awarded to Alan L. Montgomery of
the University of Pennsylvania and Peter E. Rossi of the University
of Chicago for their paper "Estimating Price Elasticities with
Theory-based Priors", to appear in the Journal of Marketing Research.
The announcement of the award was made at the August 1999 Joint
Statistical Meetings in Baltimore. Members of this year's prize
selection committee were Gary Koop, Mike West, and Max Morris
93 Mitchell Prize 2000: Announcement and solicitation.
The Mitchell Prize is awarded in recognition of an outstanding paper that describes how a Bayesian analysis has solved an important applied problem. The 2000 Prize includes an award of $1000 and a commemorative plaque, and will be announced and presented at the ISBA 2000 meeting in Crete (May 28-June 1 2000).
The Mitchell Prize is named for Toby J. Mitchell and was established by his friends and colleagues following his death from leukemia in 1993. Toby was a Senior Research Staff Member at Oak Ridge National Laboratory throughout his career, aside from leaves of absence spent at the University of Wisconsin and at the National Institute of Environmental Health Sciences. Toby won the Snedecor Award in 1978 (with co-author Bruce Turnbull), made incisive contributions to statistics, especially in biometry and engineering applications, and was a marvelous collaborator and an especially thoughtful scientist. Toby was a dedicated Bayesian, hence the focus of the prize.
This is the fourth Mitchell Prize, the first three having been awarded in 1994, 1997 and 1999. Since 1999 the Prize is awarded annually under the cosponsorship of the ASA Section on Bayesian Statistical Science (SBSS), the International Society for Bayesian Analysis (ISBA), and the Mitchell Prize Founders' Committee. The awarding of the Mitchell Prize is governed by the Mitchell Prize charter, established in 1999 (and available at the Mitchell Prize web site, noted below). Under this charter, the sponsors annually establish a selection committee; the 2000 Prize selection committee members are Gary Koop, Henry Wynn and Mike West (chair).
To be eligible for the 2000 Prize, a paper will either have appeared in
a refereed journal or refereed conference proceedings since January 1
1998, or be scheduled for future publication in a refereed outlet.
Candidate papers will be accepted from nominators and from authors. In
reviewing submissions, emphasis will be placed on evidence that the
application has truly benefited from a Bayesian analysis
ISBA Newsletter, September 1999 NEWS FROM THE WORLD
Visit the web site
to learn more about the Mitchell Prize and the sponsoring organizations.
A European Society for Industrial and Applied Statistics (ESIAS) is
being formed. A draft proposal for the mission of ESIAS is to:
foster and facilitate the use of statistics to the benefit of
European industry and other organizations; provide a forum for
networking among all users of statistics whether they are professional
statisticians, practitioners or interested users;
provide mechanisms for the professional development of statistical
practitioners and applied statisticians in Europe;
nurture the friendly and collegial interaction among and professional
development of statistical practitioners.
A list of a few representatives who will be committed to develop the
network in their respective countries is under construction.
See www.ibisuva.nl/ESIAS for more information.
Bayesian mortality models in Mexico (by Manuel Mendoza-Ramírez, CNSF). As in many other countries, in Mexico the local insurance industry uses of a number of mortality tables to calculate the premiums and reserves associated to life insurance policies. In particular, the most widely used tables, EMI82-89 and EMG73-83 (produced nine and fifteen years ago respectively), use data provided by the insurance companies. The basic procedure to obtain a table like these includes an artificial, most of the times arbitrary, increase of the observed death rates - as a protective measure - and a deterministic fit of a curve - the graduated table - describing the relationship between death probability and age.
Even though the insurance industry (as well as the government agency in charge of its supervision, the CNSF - Comisión Nacional de Seguros y Fianzas) is well aware of the importance of having updated tables at hand, these tools are replaced only after very long periods of time. Even worse, there is no systematic revision of the possible changes in the mortality patterns.
Under these circumstances, the CNSF has initiated a research project whose main objective is to develop a system by means of which the mortality patterns will be reviewed every year and, on this basis, adjustments to the mortality tables will be carried out.
This project also includes research on statistical models
for mortality, aiming to obtain
properly overestimated graduated tables. In this respect,
the CNSF has announced the future adoption of
the Bayesian methodology to produce the next generation of mortality tables
for the Mexican insurance industry. This statistical approach is based on
the analysis of the joint predictive distribution of the death rates to be
observed in the future, and makes use of some other notions such as the
``value at risk''. A paper describing these ideas in full detail is now in
ISBA Newsletter, September 1999
Prof. Valen Johnson, ISBA Treasurer Phone: +1-919-684-8753 Institute of Statistics and Decision Sciences Fax: +1-919-684-8594 Duke University E-mail: firstname.lastname@example.org Durham, NC 27708-0251, USA http://www.isds.duke.edu/~valen
Travel grants: We are pleased to announce that we have been able to raise funds for travel grants. The Savage Foundation, the Section on Bayesian Statistical Science of the American Statistical Association, and the European Union, through Eurostat, have generously contributed monies that will be used to partially cover travel expenses of a sizeable proportion of participants to ISBA 2000. In addition, financial help has also been requested from NSF. An announcement of the competition for travel grants will be forthcoming before the end of the year. Please note that only those who have submitted a paper to ISBA 2000 are eligible for support. Funds will be available to eligible participants from any country. Details about the competition will be given in the announcement.
|ITEM||COST / UNIT||QUANTITY||TOTAL|
|Single occupancy, ISBA member||$795|
|Single occupancy, non-ISBA member||$820|
|Single occupancy, student ISBA member||$695|
|Single occupancy, student non-ISBA member||$720|
|Double occupancy, ISBA member||$720|
|Double occupancy, non-ISBA member||$745|
|Double occupancy, student member||$625|
|Double occupancy, student non-ISBA member||$650|
|One child under 12, same room as parents||free|
|Additional children under 12, different room||$325 each|
|Additional hotel nights Single||$90/day|
|Full-board complement||$15 /day|