THE LONDON SCHOOL OF HYGIENE & TROPICAL MEDICINE
DEPARTMENT OF EPIDEMIOLOGY AND POPULATION HEALTH
MEDICAL STATISTICS UNIT
PhD studentship in statistical epidemiology
We are inviting applications for a PhD studentship in statistical epidemiology, funded for 3 years, starting at a mutually convenient date in 2003. The studentship is open to UK and other citizens and covers tuition fees and a generous living allowance.
Possible projects include (i) statistical methods in pharmaco-epidemiology using large general practice
databases (ii) statistical evaluation of propensity score and covariate adjustment methods for analysis of non-randomised treatment comparisons, and (iii) development of statistical methods for handling missing data in epidemiological studies and clinical trials.
Applicants must have an MSc (or equivalent) in statistics. Some knowledge of epidemiology is desirable but not essential: an aptitude for applied methodological research in statistics is more important.
Applications, including a CV and the names of two referees, should be sent to Dr James Carpenter, Medical Statistics Unit, London School of Hygiene & Tropical Medicine, Keppel Street, London WC1E 7HT (email [log in to unmask]), from whom further particulars can be obtained. For an informal discussion telephone James (020 7927 2033) or Professor Stuart Pocock (020 7927 2413). The closing date for applications is Friday 31st January 2003.
************FURTHER PARTICULARS***********************
Please find below:
1. Further information about the School
2. Further information about the projects
1. FURTHER INFORMATION ABOUT THE SCHOOL
The London School of Hygiene & Tropical Medicine has restructured for the year 2000 and beyond. A rapid process of adaptation and change has confirmed the School as Britain's national school of public health and has prepared the School for an enhanced role in Europe, while retaining its established world-wide reputation in tropical medicine and tropical public health. Both European and Third World concerns are underlined by the School's expertise in preventive medicine, epidemiology and disease control. The School currently employs about 560 staff, and there are 500 full-time equivalent postgraduate students following taught courses or undertaking research training. More details are available on the School's World-wide Web site at http://www.lshtm.ac.uk
The Department of Epidemiology & Population Health
This houses the largest group of epidemiologists, statisticians and medical demographers in Europe, together with nutritionists, social scientists and public health practitioners, working on the diseases of major public health importance in both the industrialised and the less developed countries. EPH has approximately 150 staff members organised into the following six research units: Cancer and Public Health Unit, Centre for Population Studies, Epidemiology Unit, Maternal and Child Epidemiology Unit, Medical Statistics Unit and Public Health Nutrition Unit.
The Department has a teaching programme consisting of seven MSc courses: Communicable Disease
Epidemiology, Epidemiology, Medical Demography, Medical Statistics, Public Health, Nutrition, Reproductive & Sexual Health Research, and Public Health in Developing Countries (run jointly with the School's Department of Public Health and Policy and Department of Infectious and Tropical Diseases). The Department also has approximately 50 higher degree research students studying for an MPhil, PhD or DrPH degree. The Department Head is Professor Charles Normand.
Medical Statistics Unit
The Medical Statistics Unit specialises in methodological research in medical statistics, especially in relation to longitudinal data, missing data, clinical trials, observational epidemiology and disease prevention. The Unit incorporates a Clinical Trials Research Group (concerned with planning, co-ordination, statistical analysis and reporting of clinical trials), and has a special interest in cardiovascular disease, asthma, HIV and perinatal studies. The Unit has established a reputation for being one of the leading innovative centres in Europe for biostatistical methodology relevant to the planning and reporting of medical research. This position was further strengthened by the recent appointments of Professor Mike Kenward to the GlaxoSmithKline chair in biostatistics and Professor Stephen Evans to the chair of pharmaco-epidemiology. The Unit Head is Professor Diana Elbourne. Further details can be found on the Unit's web site, http://www.lshtm.ac.uk/msu
2. FURTHER INFORMATION ABOUT THE PROJECTS
PhD projects in medical statistics aim to develop and refine statistical methodology applicable to medical
problems. Three project proposals are given below, but others are possible. Prospective applicants should consult the Unit's web pages (http://lshtm.ac.uk/msu) for further details about Unit research interests, and telephone James Carpenter (+44 (0) 20 7927 2033) or Stuart Pocock (+44 (0) 20 7927 2413) for an informal discussion.
Project proposal 1: Statistical methods in pharmaco-epidemiology using large general practice databases
The project concerns design issues and statistical methods in pharmaco-epidemiology, i.e. the study of the use and effects of drug treatment in populations, both as regards effectiveness and safety. Although this can be viewed as an application of epidemiologic methods to pharmaceuticals, the ways in which pharmaceuticals are prescribed, employed, marketed and regulated markedly increase the risk of bias in observational analyses. Thus this field poses challenges that often require special solutions in both study design and statistical analysis. Taken together with the large databases which are now available, such as the General Practice Research Database, the complex and dynamic context of pharmaco-epidemiology gives rise to fascinating and unique statistical challenges.
For instance, for any particular class of drugs (eg statins for reducing risk of coronary disease) it would be appropriate to investigate possible associations with several other diseases (eg Alzheimer's disease, eye cataract, suicide etc). The case-cohort design seems well suited to such problems by making use of all disease cases together with a random sub-sample of the whole cohort. Nested case-control studies for each disease might be a suitable alternative approach. We intend to pursue a program of methodological research aimed at determining the best approaches to such pharmaco-epidemiological study designs and statistical analyses. In order to enhance the practical relevance of this research it will be closely linked to and illustrated by specific potential drug-disease associations of interest to pharmaco-epidemiologists.
The precise sequence of methodological issues to be tackled will unfold over time but specific potential topics of interest are as follows:
1. One problem with the case-cohort approach is the difficulty in taking account of general practice effects since there is no matching of cases to controls. How much does this apparent deficiency matter?
2. It is unclear what size of sub-cohort should be selected and how one should restrict its sampling frame to take account of the characteristics of disease cases;
3. It has been suggested that repeat use of the same sub-cohort for multiple disease-drug inferences may have some statistical penalty re non-independence;
4. The methods of statistical analysis for case-cohort designs are still being developed, and experience is needed inical Statistics Unitn their practical implementation in large-scale applications;
5. The use of nested case control studies has the inefficiency of needing a separate selection of matched controls for each set of disease cases. However, one can easily match on practice and then the statistical analysis methods are better established. It will be valuable to compare results (and the effort to produce them) for nested case-control and case-cohort design for the same drug-disease questions;
6. Given the novel approaches required, the development of appropriate statistical software and experience in handling such large databases will be useful for future applications.
Our methodological findings, backed up by experience in real-world pharmaco-epidemiological questions
should have a major impact on how best to make use of such large databases as the GRPD for studying drug safety and effectiveness in a primary care setting.
Project proposal 2: Statistical evaluation of propensity score and covariate adjustment methods for analysis of non-randomised treatment comparisons.
For non-randomised studies comparing alternative treatments, it is important to take account of known
differences in the selection of patients onto the two or more treatments, as measured by covariates recorded at patient entry into the study. Ignoring such selection bias could seriously distort the comparative findings. Also, in randomised trials it is common practice to undertake covariate adjusted analyses to obtain the most appropriate and statistically efficient estimates of treatment differences, which correct for any chance baseline imbalances in covariates of prognostic importance.
There exist standard methods for taking account of baseline differences. The outcome is modelled with
treatment as a dummy variable, while covariates are included in model according to (i) their intrinsic interest; (ii) whether they predict outcome, and (iii) whether they are unbalanced between treatment groups. However, there are various problems. For example, should all covariates be included or just a relevant subset? What statistical algorithm should be used to select covariates? What is the relative importance of (i)-(iii) above? How should departures from linearity and interactions be handled? How can we avoid miss-specifying the model, which might introduce bias?
An alternative strategy in non-randomised studies is to use a propensity score. Briefly, on the basis of all known covariates one undertakes a discriminant analysis with treatment group as the binary outcome (in so doing the response is not used at all). Then the distribution of the propensity score is used to define a number of strata, and a stratified analysis of the outcome by treatment group is performed. Again, there are a number of potential problems. For example, how should the statistical algorithm for determining the propensity score be defined? How should the number of strata derived from the propensity score be optimally defined? Propensity scores take no account of covariates that predict outcome, so a feature of prognostic importance well matched between the groups may not be included. How should this be handled? How can the method be extended to more than two treatments?
The proposed project will compare these methods, and address these questions, using a variety of data sets available at the School, including the General Practice Research Database. The overall goal is to come up with clear guidelines on how covariate adjustment and propensity scores should and should not be used in reports of non-randomised and randomised studies of treatments, with a particular emphasis on well specified statistical analysis plans for regulatory submissions.
Project proposal 3: Development of statistical methods for handling missing data in epidemiological studies
and clinical trials
It is very common in research studies for a proportion of intended observations to be missinnical Statistics Unitg. This proportion inevitably rises if the study is longitudinal. If the proportion of missing observations is non-negligible, there exists uncertainty in the inferences drawn that goes beyond familiar statistical imprecision.
Such missing data problems have received great attention in the statistical literature. While appropriate methods for basic cross-sectional data are now fairly well established, advances in statistical methodology for longitudinal and more general multi-level data have led to much recent research into the issues surrounding missing data in these contexts. Compared with the basic cross-sectional case, the issues are considerably more complex and the methodology is both less developed and less accessible to researchers.
It is appropriate to analyse incomplete data using the statistical methods one would have used if the data were missing by design if certain probabilistic assumptions about: (i) the probability of the outcome being missing; (ii) the value the missing observation would have taken; and (iii) the data actually observed, are valid. These assumptions are collectively known as missing at random (MAR). Unfortunately, though, the validity of MAR assumptions cannot be determined from the data.
In fact, it is only in exceptional circumstances that MAR assumptions can be justified. Thus, while some have argued that the corrections implied by a valid MAR analysis (performed, for example, using multiple imputation) are sufficient, MAR models are a restrictive, and not particularly plausible, class of models in the presence of missing data.
This therefore points to the use of sensitivity analyses, where, in the presence of missing data, the sensitivity of the scientific conclusions is examined for robustness to MAR assumptions. The problem is that the approaches commonly adopted for sensitivity analysis, such as last observation carried forward, are so ad-hoc that it is usually difficult to attach scientific meaning to the envelope of results produced. A more sophisticated approach is therefore required: we identify three broad strands in the recent literature.
The first is parametric model based. Here, the robustness of conclusions derived from the MAR model is
assessed by nesting it within a non-MAR model. The second approach involves evaluating the complete set of inferences compatible with the observed data. This is particularly attractive for discrete data, and leads to the identification of a region of the likelihood (and hence a set of parameter values) compatible with the observed data. The third, technically demanding, approach has been pioneered by Robbins, Rotnizky and co-workers and uses weighted estimating equations.
This project will review the methods currently available in the literature, and assess their potential for use in routine analysis of longitudinal follow up of clinical trials where data are missing. The most promising methods will be evaluated on a range of study data available at the London School of Hygiene and Tropical Medicine. Short comings will be identified, and proposals for refining the methods developed.
Dr James Carpenter
Lecturer in Medical Statistics
Medical Statistics Unit
London School of Hygiene and Tropical Medicine
Keppel Street, London, WC1E 7HT, United Kingdom
Tel: +44 (0) 20 7927 2033 - direct
+44 (0) 20 7927 2230 - secretary
+44 (0) 20 7637 2853 - fax
WWW: http://www.lshtm.ac.uk/eph/msu
|
