2004-2005 Session
October 13th 2004, 2pm to 5pm at MANDEC (Manchester Dental Education Centre), Higher Cambridge Street (building 43, entrance on corner facing building 37) Joint meeting with
Manchester University's Biostats Group Theme: "Problems in Survival Analysis" MILENA FALCARO and ANDREW PICKLES Univeristy of Manchester Analysis of multivariate survival data with complex structure and interval censoring The multivariate normal model provides great flexibility for complex model specification through the approach of covariance structure analysis. We describe how the 3-category multivariate ordinal probit framework can be used in this context and made more appropriate to survival data by the joint estimation of a Box-Cox transformation of the time-scale. We illustrate the method by the joint analysis of 4 survival times, censored ages of onset for two measures where both measures are recorded for each twin in a twin-pair, in order to assess the correlation in genetic and environmental effects that influence each measure. ROBIN HENDERSON University of Lancaster Joint Modelling of Longitudinal and Event Time Data Many scientific investigations generate both longitudinal measurement data, with repeated measurements of a response variable at a number of time points, and event history data, in which times to recurrent or terminating events are recorded. This talk provides an overview of methods and models for the joint behaviour of such data, concentrating on the most common situation of Gaussian longitudinal data and proportional hazard/intensity event time data. Modelling, estimation and diagnostics are discussed and illustrated. JONATHAN STERNE (Bristol) Use of marginal structural models to estimate causal effects in longitudinal studies Use of standard regression models for the analysis of cohort studies with time-updated measurements may result in biased estimates of treatment effects if time-dependent confounders affected by prior treatment are present. A covariate is a time-dependent confounder if it predicts future treatment, and future outcome, conditional on past treatment. If, additionally, past treatment predicts current covariate value (e.g. if the covariate is on the causal pathway between treatment and the outcome) then standard survival analyses with time-updated treatment effects will give biased treatment effect estimates. Marginal structural models address this problem, and can thus be used to make causal inferences about the effect of treatments in longitudinal studies. I will illustrate the use of marginal structural models to estimate the effect of highly active antiretroviral therapy (HAART) in the Swiss HIV Cohort Study. December 15th 2004 at MMU, 4.30pm for 5.00pm DAN GROVE Teaching statistics to engineers Many engineers in industry learn statistics, often as Six-Sigma, but decreasingly at undergraduate level and its status among academic engineers is low. The situation will not improve unless we change what we teach and how we teach it. We discuss pointers to the "how" from industrial courses that adopt a process-driven approach. Dan Grove has been an independent consultant in Statistical Engineering since 1989, majoring in designed experiments and statistical modelling applied to robust product design and development. He has trained and coached numerous internal consultant/trainers, including Six-Sigma Black Belts and Master Black Belts, for Ford Motor Company, Jaguar Cars and Perkins Technology. With Tim Davis of Ford, he wrote the book Engineering, Quality And Experimental Design (Longman, 1992) which was a pioneering attempt to present statistical methods within an engineering framework (rather than a statistical textbook with engineering examples). He has developed or helped to develop a large amount of training material, including (from 1990) sections of Ford Europe's Engineering Quality Improvement Program and (from 2000) a course in Statistical Engineering (SE) for Ford's world-wide Technical Education Program. The SE material introduces many of the statistical methods needed in robust design and Design for Six-Sigma. February 9th 2005 at MMU, 4.30 for 5.00pm DAVE COLLETT (UK Transplant, Bristol) Some Statistical Problems in Organ Donation and Transplantation There are many important problems stemming from organ donation and transplantation where statistical input is needed for their solution. These include audits to identify the potential for organ donation in the UK, studies designed to improve the efficiency and equity of organ allocation procedures, monitoring the performance of individual transplant units to identify any whose results become out of line, and analyses of outcomes following organ transplantation. In this talk, some specific topics in these areas will be described and illustrated, with reference to the application of multilevel logistic modelling, frailty models for survival analysis, risk adjusted cusum analyses, and models that incorporate time dependent variables. March 16th 2005 at MMU, 4.30 for 5.00pm PAUL CLARKE (Department of Infectious Disease Epidemiology, Imperial College London) New predictions of the British vCJD epidemic [This meeting is part of the Royal Statistical Society's contribution to National Science Week 2005, a nationwide programme of scientific events and activities. The Press Release can be seen here.] Variant Creutzeldt-Jakob disease (vCJD) was first identified as a new disease in 1996 and is thought to be caused by the same aetiological agent behind bovine spongiform encephalopathy in cattle. There have been 147 deaths from vCJD in all, but the number of new deaths has been declining steadily since the peak of 28 cases in 2000 to 8 in 2004. However, recently published results estimate the prevalence of vCJD infection to be 235 per million, a much higher figure than had been expected. These results challenge some of the key assumptions about vCJD used in mathematical and statistical models for predicting the size of epidemic. In particular, it has been assumed that everyone infected will eventually develop clinical vCJD and die, and that only those people who are methionine homozygous at codon 129 of the prion protein gene are susceptible. In this presentation, I shall talk about how we extended previous models for the vCJD epidemic to relax these assumptions, and the implication this has for predictions of the epidemic size. I shall also talk about the possibility of a secondary vCJD epidemic via blood transfusion and other routes, and introduce mathematical models to predict the size of such an epidemic. Some references:
April 20th 2005 at MMU, 4.30 for 5.00pm IAN WHITE (MRC Biostats, Cambridge) Randomisation-based efficacy estimators in randomised trials: when are they useful? In randomised trials with departures from allocated treatment, it may be desirable to estimate the causal effect of treatment itself. Intention-to-treat analysis does not do this, although it provides an unbiased comparison of treatment policies as implemented. Per-protocol analysis is the most commonly used alternative in clinical trials, but it is subject to selection bias. I will describe randomisation-based efficacy estimators, which avoid both these problems, and give some examples where they have led to useful conclusions. 18th May 2005 at MMU, 4.30 for 5.00pm(preceeded by a short AGM) VLADIMIR VAPNIK (Royal Holloway, London) Problem of empirical inference in machine learning and Philosophy of Science 1. The problem of Empirical Inference is a core problem of human intelligence. It has been discussed for more than 2,000 years. However only since the 1960s with the appearance of fast computers did it become a full-edged subject of Natural Science (as Physics or Biology). Now one can not only speculate about models of learning, but can also conduct wide scale experiments with computers. Such experiments have demonstrated that many prejudices were accumulated during the time when Empirical Inference Science was driven mostly by speculation. Since the 1970s Machine Learning has made great progress. In particular, complete answers to core mathematical problems of generalisation were found. 2. In my talk I will discuss results of the Mathematical Learning Theory from general point of view of Philosophy of Science. I will restrict myself mostly to facts that have the status of necessary and sufficient conditions and, therefore, must be satisfied by any learning system (including humans). I will try to show the restrictiveness of understanding learning problems in humanist studies and in particular in classical Philosophy of Science. 3. A crucial point in Machine Learning Science was the discovery (both in theory and in experimental studies) of the existence of direct non-inductive methods of inference (from data to data, avoiding a generalization stage). It has been proven that non-inductive inferences are always more accurate than inductive. It has been shown that there are exist a large family of non-inductive methods. 4. The problem of Empirical inference - along with pure mathematical concepts - contains concepts that have clear humanist interpretations (for example "cultural Universum") that are important instruments for effective inferences. I will try to show that the problem of Empirical Inference has reached the point where progress requires the joint efforts of philosophers and statisticians. In particular, my main thesis will be that any breakthrough in understanding learning will require primarily philosophical (conceptual) rather than mathematical (technical) advances. |