RA5a: Structure,environment and staffing policy
[Note: Staff returned as research active under UoA24 are in bold. Other past and present members of the Department are in bold italic. Research Assistants, as listed in RA6d, are further indicated by (a), and Research Students by (s)]
University College London Department of Statistical Science conducts basic and applied statistical research on a very broad front, with a strong interdisciplinary and collaborative focus, and with extensive development of fundamental philosophy, theory and methodology. A wide range of individual and group research activities is held together by a shared belief in the unity of Statistics, and in the importance of two-way interaction between theory and general methodology on the one hand, and the scientific requirements of real-world applications on the other. This philosophy guides the selection and directions of our investigations, generates fruitful interchange between researchers with different specialisms, and has generated significant novel contributions both to Statistics and to Science. We are committed to publishing innovative methodological work in subject-matter outlets, in order to maximise its scientific impact.
For internal management purposes, the Clinical Operational Research Unit was transferred to Mathematics (UoA 23) in 1997, producing large reductions in total research funding and research staff figures for Statistics, but without affecting academic interactions or other research activities.
ARRANGEMENTS FOR PROMOTING/SUPPORTING RESEARCH. Staff are encouraged to pursue and develop their individual research interests. The Department provides study, library and computing facilities and organises weekly seminars with internal and external speakers. A Research Report series, with internal refereeing, encourages early critical exposure and circulation of ideas: most reports are later published. The Departmental Computer Administrator ensures availability and smooth running of requisite hardware and software. Attendance by staff and students at conferences, scientific meetings and other external professional activities is given high priority and support, with funding from the Department and the UCL Graduate School. The Departmental Research Coordinator and Research Committee encourage and oversee research, and, with the UCL Research Funding Unit, ensure that researchers are aware of funding and other opportunities. We greatly value interchange of ideas with academic visitors: recent long-term visitors included B Clarke (UBC), M Gomez Villegas (Madrid), I Helland (Oslo), D Matthews (Waterloo), A Takemura (Tokyo).
COLLABORATIVE RESEARCH. A high proportion of our research involves intensive long-term collaborations with applied subject-matter specialists. This is crucial to our mission to identify, develop and apply the sound statistical foundations essential for scientific advance. Major collaborative research areas include Statistical Hydrology; Climate Research; Mathematical Biology; Forensic Identification; Human Genetics; Medical Statistics; Pharmaceutical Statistics; Complex Instrumentation; Geological Dating. Extensive consultancy for external bodies, including Environment Agency, Meteorological Office and many pharmaceutical and insurance companies, has given further impetus and direction to our research. Amgen is funding Rosati (a) to work on optimal project scheduling. Student support has come from Glaxo SmithKline, Novartis, and the Italian National Institute for Food and Nutrition Research. Two students were based at ICRF. The rich and diverse environment of University College London generates many exciting research links, in particular with UCL interdisciplinary research units: CoMPLEX — Centre for Mathematics and Physics in the Life Sciences and Experimental Biology (Bioinformatics); Gatsby Unit for Computational Neuroscience (Probabilistic Expert Systems, Machine Learning); Institute for the Natural Environment (Climate Modelling); Benfield Greig Hazard Research Centre (Geohazards).
STAFFING POLICY. We have a long-standing policy of appointing young staff of high promise. Wilkinson-Herbots (Lecturer 1999– ) was Royal Society Dorothy Hodgkin Research Fellow (UCL, 1995–9). Hazelton (Lecturer 1994–7) is now Senior Lecturer, University of Western Australia, and Vice-President, Statistical Society of Australia (WA Branch). In the assessment period he published 10 refereed articles. de Luna (Temporary Lecturer 1997–9) is Research Fellow in Economics, Umeå University, Sweden, with 6 refereed publications. Pauler (Temporary Lecturer 1996–7) is Statistician, Fred Hutchinson Cancer Research Center, Seattle, with 6 refereed publications. Skouras (Lecturer to June 2000) has taken a research position in the City, and his UCL appointment has been transferred to Honorary Senior Research Fellow. He continues to lecture and supervise research students, contributes actively to collaborative research, and is a member of our EU TMR group. Chandler was a Research Assistant (joint with Civil and Environmental Engineering, Imperial College) until his appointment as Lecturer in June 1997. He has already developed a strong reputation in climatological modelling and analysis. Didelez (Lecturer January 2001– ) has just completed her PhD but is already achieving international recognition. She is actively developing several fruitful lines of theoretical and applied research, both individual and collaborative (see “Research Activities” below and RA5c). Of her RA2 submissions one paper has already appeared in Biometrika, another (report 186) is accepted for Biometrical Journal, and a third (report 110) is provisionally accepted for Statistica Neerlandica. Another recent appointment is Downie (see RA6c). Probationers are assigned a senior colleague as mentor, and allocated light teaching and administrative loads. The UCL Review and Development scheme enhances the performance of both new and established staff.
We benefit from the active involvement of retired and honorary staff. A paper by Emeritus Professor Stone (with Hazelton) was read at a RSS Ordinary Meeting and published in JRSS(A). Stone also organised and spoke at a RSS Official Statistics meeting. Honorary Professor Royston (MRC Clinical Trials Unit), who in 1996–2000 published 35 methodological and application papers, is collaborating with Omar on model validation and assessment of model uncertainty, and with Farewell on longitudinal studies. Visiting Professor Jones (Glaxo SmithKline) is supervising a PhD student, and working with Senn on individual bioequivalence, multicentre clinical trials and cross-over designs, as well as on genetic experimentation.
Our belief that applied statistical researchers should maintain an active interest in general methodological development is reflected in several joint appointments. Senn (Professor) and Clarke (a) (Research Statistician) are joint with Epidemiology and Public Health. Omar (Senior Lecturer) is joint with UCLH NHS Trust. Copas (Lecturer 1998– ) is joint with Sexually Transmitted Diseases, and is being returned as research active under UoA 3a. Wright (Temporary Lecturer 1996–9) was partially supported by Great Ormond Street Hospital for Sick Children.
RESEARCH TRAINING. The Director of Research and Departmental Research Committee oversee admission and progress of research students. Research studentships, including CASE awards, are obtained from EPSRC, MRC, and BBSRC. Weekly supervision is the norm. The UCL Graduate School lays out what is expected of students and supervisors in its Code of Practice, and is piloting a student log book. We operate a training and monitoring programme with six-monthly milestones and reports, including seminar presentations. Convincing evidence of ability, including a substantial written report, is required before upgrading from MPhil to PhD. Every student is assigned a second supervisor, who plays an active role in assisting and assessing progress. Students are encouraged to attend Graduate School training courses in generic and transferable skills, to which this Department itself contributes a course on Statistics for research students in experimental sciences and other disciplines. Our students may be required to attend relevant modules of our MSc course, including “Communication Workshop”.
Overall the plans and expectations of our RAE 1996 submission have been accomplished and surpassed. In addition, there has been extensive achievement in new research topics and areas.
FOUNDATIONS AND THEORY OF STATISTICAL INFERENCE. We aim to clarify and extend the logical foundations of Statistics and explore their implications for applied data analysis. Since RAE 1996 Dawid has developed and extended his theory of Prequential Analysis, a unifying framework for Statistics centred on the production and assessment of a stream of probability forecasts. Its new foundation in sequential game theory has led to fruitful links with other active research areas, including Computational Learning Theory and Economics. Game theory also underlies a novel approach by Dawid and Grunwald (Eindhoven) to Maximum Entropy, which has been generalised to arbitrary decision problems, and shown to be equivalent to robust Bayes analysis. Dawid and Vovk (Royal Holloway) developed new philosophical and mathematical foundations for probability theory within the prequential setting. Dawid and Skouras studied prequential variants of “consistency” and “efficiency”, obtaining powerful general results without requiring independence, ergodicity or well-specified model. Skouras and de Luna constructed prequential methods to guide data-dependent choice between model-selection criteria. Dawid's contributions to Probabilistic Expert Systems (PES) were published in a joint monograph and his algorithms incorporated into commercial software (Hugin). New work focused on learning and criticising qualitative and quantitative structure. In Bayesian Inference, Dawid and Sebastiani (Amherst) identified close relationships between Bayesian and classical approaches to optimal experimental design. Dawid and Lauritzen (Aalborg) studied methods for assigning compatible prior distributions to multiple models, identifying and eliminating ambiguities in existing suggestions. In Likelihood Inference, Farewell and Cook (Waterloo) developed a new mixed form likelihood for combining different data types. Farewell and Sprott (Waterloo) developed conditional likelihood methods for predictive agreement.
Dawid's theory of Conditional Independence was the subject of a 1999 international workshop at the Fields Institute (Toronto). He has recently transformed it into an abstract concept of irrelevance based on a novel algebraic system, the “separoid”, with applications to a wide range of uncertainty formalisms in Statistics and AI.
Dawid developed a novel probability-based decision-theoretic approach to Causal Modelling and Inference, eschewing counterfactuals. Didelez constructed new graphical representations of causal relations between streams of events, together with graphical semantics for manipulating the associated non-symmetric local independence property. These simplify causal inference by clearly displaying collapsibility, likelihood factorisation, etc. Didelez and Edwards (NOVO Nordisk, Denmark) have identified necessary and sufficient conditions for collapsibility in CG-regression models, so streamlining the analysis of complex mixed discrete-continuous variable problems through decomposition into smaller and more manageable parts.
APPLIED PROBABILITY AND STOCHASTIC MODELLING. Since RAE 1996 Isham and Chandler have developed Modelling and Inference for Spatio-Temporal Processes. Their research and related consultancy in Hydrological and Climate Modelling and Analysis, with an extensive network of collaborators, PDRAs and students, is widely recognised by practitioners as improving substantially on existing methodologies, for example for flood risk assessment. Continuous space-time stochastic point process models of precipitation, developed by Isham with Cox (Oxford) and Rodriguez-Iturbe (Civil Engineering, Princeton), were extended by Isham, Chandler, Northrop (a), Onof and Wheater (Civil Engineering, Imperial College) to provide realistic models of fine-scale structure and to support continuous simulation, and were extensively validated against high-resolution data. New moment-based inferential methods were compared with Chandler‘s Fourier-based approximate likelihood method. Criteria of model fit using scaling and extreme-value properties were developed and implemented. Isham, Cox and Rodriguez-Iturbe constructed a realistic stochastic model of soil-moisture temporal dynamics incorporating various modes of input and loss, enabling investigation of various environmental scenarios. Chandler and Yan (a) (ex Institute of Atmospheric Physics, Beijing) constructed Generalized Linear Models for climate data to realistically represent uncertainty, non-stationarity (including global warming) and atmospheric dynamics. These models are being applied to risk assessment in the UK insurance industry. With collaborators in Space and Climate Physics, Chandler produced the first probabilistic forecasts for seasonal tropical cyclones. Chandler developed a Bayesian method for image analysis to infer a fine-scale precipitation field (as required for hydrological use) from observed coarse-scale data and other relevant information. This flexible new methodology is under trial by UK Meteorological Office for inclusion into their operational forecasting system.
In Mathematical Biology, Isham and Mollison (Heriot-Watt) organised an EPSRC/RSS research workshop on Stochastic Modelling and Statistical Analysis of Epidemics. Stochastic modelling of Host-Parasite Systems yielded novel understandings with important practical implications. Mathematical models of macroparasitic infections, developed by Isham and Herbert (s), led to improved insight into mechanisms affecting overdispersion of parasite counts. Isham, Cornell (a) and Grenfell (Zoology, Cambridge) found that invasion by drug-resistant macroparasites is increased by spatial aggregation. Non-linear effects produce counter-intuitive behaviour: thus epidemics, when rare, may be more likely and more severe in small host populations. Isham and Chan (Zoology, Oxford) evaluated rival explanations for variability of parasite burdens, and studied the extent to which immune responses can be inferred from patterns of infection. Isham, Webberley (a) and Hurst (UCL Biology) are developing realistic models for the complex process of infection of coccinellid beetles by sexually transmitted mites. By modelling transmission dynamics, Isham and Xiridou (s) were able to compare alternative TB control strategies.
In Mathematical Genetics Wilkinson-Herbots developed Markov Chain models for genealogy and genetic differentiation, showing that asymmetric population structure substantially increases differentiation between subpopulations and its dependence on mutation. In Human Genetics she identified a mitochondrial (mtDNA) polymorphism holding significant information about the origin and history of European populations. This enhanced and clarified interpretation of a European mtDNA database. Her analysis with Richards and Sykes (Oxford) and Forster (Hamburg) of UK and European mtDNA identified major genetic clusters and suggested plausible novel hypotheses about their origins. A Chinese mtDNA dataset is currently being studied with Decorte (Leuven).
Dawid (with Mortera, Rome) clarified several subtle logical issues in Forensic Identification, including a current controversy on how to allow for the fact that a suspect has been identified by a computer database search. He reconciled several seemingly incompatible models of genetic substructure, and with Pueschel (s) corrected existing analyses of the effect of substructure on the impact of DNA identification evidence. With Wilkinson-Herbots he explored the use of mtDNA profiling for forensic identification. In an international collaboration with statisticians and forensic geneticists, he is developing a PES-based approach to complex forensic DNA identification queries with missing individuals and mutation. This has been the subject of two “research kitchens”, funded by the European Science Foundation and Danish Research Councils.
SCIENTIFIC AND INDUSTRIAL APPLICATIONS. Fearn's research (with Brown, Kent) into variable selection for regression and discrimination with large numbers of predictor variables has resulted in new Multivariate Bayesian Methodology. Originally developed for spectroscopy where Fearn has many continuing applied collaborations, this is proving important in other actively developing areas including gene-chip technology. Methodology developed by Fearn and Gilg (s) for modelling animal growth curves is in current industrial use. Fearn‘s decision-theoretic interpretation (with Thompson, Chemistry, Birkbeck) of Fitness for Purpose in Analytical Measurement has aroused much interest in the analytical chemistry community.
Detailed statistical modelling for Fission Track Thermochronology (Galbraith with Laslett, CSIRO, and geologists in UK, USA and Australia) has had a major influence on practitioners and research scientists. Achievements since RAE 1996 include a new method of modelling fission track annealing (central to the estimation and interpretation of geological thermal histories); a method for estimating the youngest age (important in sediment provenance studies); and a fundamental study of the information contained in different types of fission track measurements. Galbraith (with Roberts, ARC Fellow, Melbourne) developed the first rigorous statistical models and methods for analysing Optically Stimulated Luminescence (OSL) data, leading to deeper understanding and improved laboratory protocols. Application areas include archaeology and natural history. Their OSL dating of single quartz grains from the important Jinmium archaeological site, leading to fundamental revision of its estimated age, was widely reported.
MEDICAL STATISTICS. Farewell‘s research (with Gladman, Toronto, and Isenberg, UCL) into statistics for rheumatology has led to novel general clinical trial methodology, including probabilistic structures for the multiple outcome measures now required in many trial guidelines. A wide-ranging review article was commissioned as an editorial. Farewell introduced a new paradigm for defining measures of responsiveness to treatment; his novel experimental designs for identifying reproducible measures of response to therapy are being implemented by international teams of investigators in lupus and myositis. Farewell‘s methodological work on lost-to-follow-up studies has become particularly important with increasing use of longitudinal medical databases. Farewell has redirected his work on modelling the UK AIDS epidemic to provide estimates of the demand on the Health Service under various scenarios for long-term treatment effectiveness and uptake. Methods developed by Farewell and Copas to incorporate retrospective data into cohort studies and for informative non-response in surveys have proved crucial to the interpretation of national surveys of sexual lifestyle used to underpin public health initiatives. Farewell and Clarke (a) are developing new methodology for non-response, with application to the major MRC/NIH Whitehall study. Of two invited presentations to the Royal Society of Medicine, one was based on Farewell‘s work (with Steiner and Cook, Waterloo) on monitoring surgical outcomes, adapting industrial cusum techniques to allow for patient heterogeneity and multiple outcomes; the second reported progress on assessment of surgical excellence (Farewell with DeLeval and Carthey, Great Ormond Street), including analysis of the occurrence of major and minor negative events during surgical procedures.
Senn contributes significantly to sound use of Statistics in the Pharmaceutical Industry, with innovative work on cross-over trials, bioequivalence, meta-analysis, baseline adjustment and strategies for drug development, and frequent lectures to the industry aimed at disseminating good statistical practice. Senn and Lambrou (s) pointed up a hitherto unappreciated distinction between two types of design robustness, and demonstrated that realistic pharmocokinetic models have implications for the design of cross-over trials very different from current practice. Contrasting Bayesian, Neyman-Pearson and standard frequentist approaches to bioequivalence testing, Senn demonstrated the dangers of slavish reliance on narrow “optimality” criteria. Senn and Grieve (Pfizer, Sandwich) showed that for the AB/BA crossover design there is less difference than had been supposed between sensible frequentist and Bayesian analyses. Senn’s concern for the foundations and history of statistics has been reflected in his book Statistical Issues in Drug Development and two RSS read papers, addressing the extent to which consensus in application is possible despite foundational disagreements.
Omar‘s prime responsibility is to enhance the quality of research in UCL Hospitals, including developing training in research methodology. Omar has constructed and demonstrated the value of hierarchical models for continuous, binary and ordinal outcomes of cluster randomised trials (with Thompson and Turner, MRC), and for meta-analysis (with Whitehead, Reading, and Goldstein, Institute of Education). With Royston, Ambler (a) and Taylor (Imperial College), Omar is devising strategies to formulate and validate risk models in cardiac surgery: a recent meeting with statisticians, cardiothoracic surgeons and health service researchers disseminated work from the first phase of this project. Omar (with Majeed and O'Sullivan, UCLH) is constructing risk adjustment models to predict use of Primary Care resources. Applied collaborations include relating breast cancer to alcohol consumption, effects of physiotherapy on knee surgery, and risk of mechanical failure of artificial heart valves.
Copyright 2002 - HEFCE, SHEFC, ELWa, DEL
Last updated 17 October 2003