Tom Leonard - The Life of a Bayesian Boy    

 
CHAPTER 7: THE UNIVERSITY OF EDINBURGH
 
Tom's former colleague Angelika van der Linde (Oldenburg, 2012)
 
"The real thing in this world is not so much where we stand, as in what direction we are moving." (Julius Nyerere)
 

The Department of Mathematics and Statistics of Edinburgh University was housed in the James Clerk Maxwell Building on the King’s Buildings campus. It consisted of the pure mathematicians, including several theoretic, self-protective probabilists, the traditional applied mathematicians, who were led in fine style by David Parker, a successful O.R. group, and several isolated statisticians who taught from lecture material that’d been prepared and meted out at the time of the Ark.

The statisticians were the residue of David Finney’s and Peter Fisk’s old Department of Statistics plus the newly-appointed double-bootstrapper Bruce Worton. He’d published some interesting research with David Hinkley, who I’d recently met with a beautiful lassie during a convivial party in John and Serene Hsu’s house in Santa Barbara.

Bruce’s position was described to me as the new appointment that came with the chair, but the highly-accomplished applied statistician Crystal Donnelly had, without my knowledge, been encouraged by the manipulative mathematicians to leave for Oxford in order to accommodate this, and her probabilist boyfriend Ben Hambly departed in a huff shortly after she left. They were presumably unaware of the full nature of the ripples from above, and I only heard the complete story later,  from an Associate Dean.

I’d inherited a tenuous political situation. The albeit well-meaning departmental chairman had extremely bizarre views on what Statistics really was, and regarded us as a mathematical example. Furthermore, a totally mysterious, apparently papier maché Government quango of ex-civil servants called BIOSS (Biomathematics and Statistics Scotland), who described themselves as a charity, was sticky-taped to our belly from the floor below. Their functions seemed, at first, sight to be to siphon off the grants in Agricultural Statistics and to award each other advanced doctorates in Statistics and honorary professorships without actually producing anything  substantial.

However, BIOSS’s director Rob Kempton, who was out of Oxford, proved to be a fine statistician and supportive colleague. He helped me to trace my great great grandfather’s (the Rev. Dr. Francis Bryant of Peter Tavy) records at Wadham College, where my accomplished ancestor was awarded a third class honours degree and a D.D., and to discover that my great great great grandfather was a gentleman in Holborn.  

Under Rob’s directorship , BIOSS proved to be immensely influential though rather all-constraining in political terms. Chris Glasbey published a series of influential papers on Bayesian Image Analysis, one of them jointly with Professor Kanti Mardia of Leeds who was one of the most brilliant statisticians in Britain.

 
   
Rob Kempton
(1946 - 2003)
  Ian Main   Alan Izenman
 
Orestis Papasouliotis and family
 

I was pleased to be able to make Bruce Worton’s appointment permanent; he was an excellent teacher, particularly on a large service course to the biologists, and I’d been intrigued by the bootstrap ever since I’d listened to Bradley Efron talking about it in Monterey in the early 1980s.  I was also delighted to be able to bring in my erstwhile research assistant Orestis Papasouliotis from Madison as the research associate in our hastily-devised STATLAB. I became its first director, and several other statisticians in the department occasionally collaborated.

STATLAB’s immense success was to later on be seriously undervalued by our Faculty Office. In hindsight, I believe that we achieved monumental heights. Bruce Worton and I started off by helping Dean Geoffrey Boulton to analyse his educational foresight data, and he extrapolated our conclusions all over the place.

Bruce did some useful research on the frequencies of bird species for RSPB, which also contributed to STATLAB’s success. I got him started with some theoretical suggestions for a Bayesian modeling problem, and he later published some business applications, with George Stretfaris, as well as publishing some applications of other methodology with BIOSS. He and Orestis later co-published their analysis of the Scottish glaucoma data. [see item (8) in my CDC section]

Orestis and I completed a major consultancy for Northumbria police by analysing aggregated patient records pertaining to a multiple murder inquiry (a nurse in Newcastle was suspected of routinely overdosing her patients).

During our complicated analysis we discovered that Bayes factors, when comparing a one-sided null hypothesis with an oppositely one-sided alternative hypothesis, can work well in frequency terms and that Lindley’s paradox can be avoided. While we came up with a number of significant conclusions, with the help of several jolly police officers visiting from Newcastle including the superintendent himself, the nurse was acquitted, possibly as a result of our analysis.

STATLAB also made substantial contributions to the template modelling of the amount of mutton in sheep, and to the Bayesian statistical analysis in Astronomy of data relating to ionising flux in high redshifts. See items (2) and (6) of our CDC section.

Orestis proceeded to develop a Bayesian approach to the analysis of covariance, as his Ph.D. topic in Edinburgh, using judicious applications of MCMC, and he applied this to a large forensic science data set contrasting different types of criminals that was provided to us by Tony Busutil, the University’s Professor of Forensic Pathology.

         The conclusions are described in some detail by Orestis in item (8) of our CDC section. They suggested that paedophiles could not be detected by neuropsychological tests, given the data then available.  I’d been previously unaware that they were published by Leonard and Papasouliotis (2002) in the Encyclopaedia of Environmetrics, and it was a pleasant surprise to discover that these contributions had been recognized in the literature.

Orestis and I also developed a Bayesian backwards prediction procedure that drew inferences about the amounts of drugs smuggled in by criminals during previous trips into the country, given the amounts that’d recently been discovered in their stomachs. I co-authored a long paper in JRSSA with Orestis et al, which described our new methodology.

         We later developed a random effects multivariate generalisation, during a congenial visit by Alan Izenman from Temple University, which accommodated measurement error when taking replicated readings of the amounts of drugs.

          Unfortunately, we discovered in hindsight that Lothian and Borders police had provided their contact person Colin Aitken with extremely precise measurements whose errors were so small that we were unable to develop a meaningful practical example! But a feather in the cap for the police. And Alan had a great time with us anyway.

As a major component of his duties for STATLAB, Orestis analysed numerous data sets from around the university, including one that was left outside his door in a brown bag by a girl with strange views of Statistics.  Highlights included an application of Laplacian approximations to data in Astrophysics that simulated new star systems, and an application of Geoff McLachlan’s package for multivariate mixtures to an analysis of the Lepenski Vir archaeological data, which reported the nitrogen and carbon content of Mesolithic and Neolithic skeletons.

          The Lepenski Vir results are briefly described on page 3 of my 1999 book with John Hsu, and we came up with a clustering that wasn’t anticipated by the  traditionally-minded archaeologists who always expected the data to divide neatly into two groups.

We later developed a modified Bayesian solution to the Fieller-Creasy problem that enabled Orestis to show, by analysing an exhaustive stream of datasets, that pronounced cytoplasmic pH gradients are not required in plant and fungal cells. We reported these results, with Parton, Read et al, in the Journal of Cell Science (1997).

Orestis extended his long list of publications when he moved to Serono International, a pharmaceuticals company in Geneva in 2001. He now has a senior appointment with them, and hopes to return to Edinburgh soon with his daughter to visit the zoo.

John Hsu visited us from UCSB on several occasions, and was appointed to an honorary fellowship in our department. In 1996, John, Li Sun, Irwin Guttman, John and I published a paper in JASA that used Laplacian approximations to marginal posterior densities for parameters of interest in normal random effects models, and also refuted the 1972 Lindley-Smith estimates (Irwin was fascinated by this, though I was bored to tears about the damned things by know). An earlier version had been well-received, with a great deal of laughter, when I presented it at the founding 1993 San Francisco meeting of ISBA.

         During 1997, two of the co-authors of the 1996 JASA paper were nevertheless irrationally attacked from a source that I’ve been asked to say was unknown, but it wasn’t Dennis Lindley and one of us suffered the consequences. While I was gratuitously insulted in the most personal terms, which seemed to make an assumption about my orientation, it’s difficult to prove cause and effect. I’d hoped that the Leonard-Lindley-Smith saga had been put to bed, but it would rear its head again briefly in 1998 at Valencia 6 before the alcohol and John Deely took over.

I’d been delighted to be able to appoint Angelika van der Linde from Bremen to a lectureship in the department on the recommendation of Bernie Silverman, and this was supposed to be a research appointment.  She’d published some seminal theoretical  results on Gaussian prior processes for regression models, and she proceeded to develop DIC with David Spiegelhalter, and powerful methods using reference priors with Jose Bernardo, who visited us later. She was later of great service to ISBA, and in particular worked as production editor for the Society’s international journal. She was also a wonderful lunch-time companion.

         Angelika was dumbstruck when the mathematicians asked her to expend a large amount of her work effort preparing model homework and exam answers, rather than spending her evenings attending the film festivals.  The undergraduates would typically wait for the model answers without doing their homework, and then complain if an exam question didn’t relate to a previously circulated solution. Many of them didn’t take lecture notes either!

[In their final year, many of the students’ scores were ratcheted up by the department’s Scrutiny Committee. This ensured that most third-class students received upper seconds and their meal tickets. We were always under strong pressure from the financially-orientated Faculty Office not to fail any students whatsoever]

In 1998, Angelika and I travelled to the Sixth Bayesian Valencia Conference together, a nostalgic return for me after nineteen years, and it was good to see Peter Freeman again. The conference was held in the Hotel Las Fuentes on the Spanish Mediterranean coast, just like the first one, though the seafood wasn’t quite as appetizing, and I enjoyed the cold water springs on the beach, where I played football with Richard Smith from North Carolina and chatted about the fundamentalist Bayesians at Warwick with Peter Lee, the colourful Provost of Wentworth College from York, while Adrian Smith, still looking remarkably young, was kissing the ladies.

         Dennis Lindley was at his charming best. In 2002, he would receive the Royal Statistical Society Guy Medal in Gold from Peter Green for his pioneering developments in Bayesian Statistics, fully twenty-six years after his retirement.

My overriding concern during the poster sessions was that far too many doctoral students were using MCMC when analysing models that would have benefited from a more direct prior-to-posterior analysis; they were therefore wasting their talents.

However, Gelfand and Smith were very much revered in this respect, and I came across as an old codger with eccentric opinions.

Steve Fienberg teased me with remorseless good humour during his after-dinner speech, and brought the house down with a series of jokes regarding my 1975 paper in Technometrics. Jim Zidek thought that I came out of it well.

         Later that evening, I played the Rev. Thomas Bayes returning from Heaven in a skit prepared by Tony O’Hagan, the Master of Ceremonies of an otherwise musical cabaret. My adventures with the wandering microphone brought the house down once again. The script has been published on the Internet by Brad Carlin, and a picture of me being interviewed by Tony can be tracked down by sourcing Pictorial Tribute: Bayesian 6 Cabaret. Life is a cabaret, my friends, life is a cabaret!

After all that, John Hsu and I finally completed our book Bayesian Methods: An Analysis for Statisticians for Interdisciplinary Research Workers (1999), the fourth advanced graduate text in the Cambridge Series in Statistical and Probabilistic Mathematics, published by Cambridge University Press.

         The book was well received by the critics, with some quibbling about its level of difficulty and its focus on the whole gamut of Bayesian marginalization procedures, including MCMC. The book has sold well (about 5000 copies), with a special edition in China, and it’s still being used by Kam Tsui at the University of Wisconsin as part of the course materials for Statistics 775. Moreover, Professor Charles Franklin used it for several years as the core text for his UW graduate course for political scientists.

Michael J. Evans of the University of Toronto wrote in Mathematical Reviews that ‘This book provides excellent up-to-date coverage of modern Bayesian statistics. The book is clearly written and at a reasonably high level. Outstanding features of the text include a number of significant applications of the methodology, the use and comparison of the AIC and BIC model selection criteria in several contexts, a somewhat unique treatment of the axiomatization of Bayesian decision theory through utility, discussion of the frequency properties of the procedures, and interesting and relevant self-study exercises that extend and deepen the content of the book.’

 
Michael Evans
 

STATLAB continued to be very successful. Orestis received BP funding from Professor Ian Main in Geophysics to work on the statistics of earthquakes and oilwells, and the three of us published three papers together in international geophysics journals. One of these was authored jointly with Kes Heffer from BP, while he was moving to Heriot-Watt, and another with two further geophysicists. We were particularly concerned about long term (i.e. long distance) correlation between earthquakes, and between oilwells.

         Our best achievement was the development of a simple algorithm for discerning the couplings of injector wells and producer wells in a large complex oilfield that might give rise to the best output. The fifth chapter of Orestis’s thesis is devoted to this topic, and contains a fascinating map of a large North Sea oilfield. His external examiner, David Wright of the University of Plymouth, said that he was most impressed most of all by this practical contribution. [I also knew Julian Stander and I’d served as external examiner for one of their students at Plymouth, as well as visiting them to give a seminar] The University of Edinburgh have since obtained international patents for our method. My older grandson was pleased to hear that I’m an inventor.

 
Julian Stander
 

Around that time, I was a co-author of three joint papers on radial basis networks in the Artificial Intelligence literature with the international chess master Mark Orr [with whom I’d won the Edinburgh team championship in 1996 with the Wandering Dragons. In 1992, I was chess champion of North-East Wisconsin]. However, as I remember, my only substantive contribution was to Mark’s EPSRC grant proposal, since one of my boy scouts (called Jet, Jip, or whatever) preferred to go walk-about. But here’s a big thank you to Mark.

Irwin Guttman visited Edinburgh from SUNY at Buffalo for several months during 1998, and he still remembered my 1975 University of London seminar on Bayesian contingency table analysis. We developed solutions to a couple of research problems together, and he and Orestis helped me to complete my book A Course in Categorical Data Analysis that I was writing under unfair time pressure from the incoming administration at Chapman and Hall. They both thought that the end product was rather good.

In 1999, Mr. John C. Duffy returned to the Statistics group, from the Scottish Health Department, after a long absence. He’d previously published some very interesting research with Timo Alanko of the University of Helsinki on the estimation of the distribution of alcohol consumption. I’d reviewed Alanko’s Ph.D. thesis on the same topic, and this helped him to publish his thesis as a book. John is currently (as of March 2012) Deputy Director of Knowledge Management for the Scottish Higher Education Funding Council, and we’ve discussed a variety of statistical ideas and all the intrigue together over the years.

         In 1999, I took the opportunity to work with John on a meta-analysis problem relating to several 2x2 contingency tables and the Mantel-Haensel model, and we developed novel Laplacian approximations to the posterior densities of the common measure of association and other parameters of interest, under fixed effects assumptions. Our methodology is published in Statistics in Medicine (2002).

         In our first practical application we combined six different studies of children with acute otitis media (an ear condition), where each study compared the proportion of patients experiencing pain 2-7 days after receiving an antibiotic treatment with the proportion in a control group who experiencing pain after 2-7 days. While the approximate posterior density of the common log measure of association was skew to the right, it showed in precise terms that pain and the antibiotic treatment were strongly negatively associated. This conclusion was further validated by a graphical variant of a L’Abbé plot. In our second practical application we investigated three different studies of children developing contralateral otitis media in a treatment group. and a control group. After investigating a variety of posterior densities, we concluded that the common measure of association assumption was not completely reasonable. In such cases, the Mantel-Haensel test should be employed with extreme caution.

 

 
 
 
 
  © Thomas Hoskyns Leonard, 2012 - 2013