Bias-adjusted inference in Biostatistics

Advancements in medical practice must be evidence based. For example, new treatments for diseases (such as lung cancer) must be proven effective before being licensed and, equally, modifiable exposures (such as smoking) must be reliably linked to adverse health outcomes before public health strategies are implemented. This necessitates the collection, analysis and interpretation of data. In today's competitive world, time and money are in short supply. Scientists are therefore under pressure to collect this evidence as efficiently as possible, by making use of existing data sources, novel trial designs and the latest technology where available. However, failure to understand the process by which data sources are created can lead to a biased picture of the information they provide, and give rise to poor medical decision making. Some examples of this now follow:

The size (i.e. the number of patients) and the scope (e.g. the question being asked) of a clinical trial has traditionally been fixed in advance. However, they are increasingly conducted in a sequential fashion, with less rigid guidelines as to the direction the trial may take. For example, patients with advanced cancer could participate in a randomised trial, but be allowed to switch to a new treatment if their initial therapy proves ineffective. Or, patients recruited in the second year of a trial testing a single therapy at multiple doses, could receive the dose which performed the best in year one. Adaptations like these mean patients can be afforded a high standard of care in the trial, but at the same time enable effective treatments to be identified and licenced more quickly. Yet, if the data arising from such trials are simply taken at face value, they can also systematically over- or under-estimate the true effect of the treatment.
One of the primary goals of Epidemiology is to find the root causes of a disease, so that it can be treated effectively. Ethical and practical reasons often mean that clinical trials - the best way of testing causal hypotheses - are not always possible. Epidemiologists then must rely on observational or retrospectively collected data to find these causes. However, strong correlations seen between a potential risk factor (e.g. alcohol intake) and a medical condition (e.g. high blood pressure) are no guarantee that alcohol causes high blood pressure, because a separate unobserved factor (e.g. salt intake) may in fact be a relate to both. This is referred to as `confounding bias'. It is difficult to adjust for the effect of confounding, because one is never sure that all possible confounders, like salt intake, have been found.

An implicit assumption made when synthesising the results of all the available medical trials addressing a particular research question, in order to inform future health policy, is that they form a representative and unbiased sample of all studies conducted. However, it is often only practically feasible to find and include studies that have been published in academic journals. When scientists cherry-pick their most exciting results to submit for publication, and the journal editors preferentially publish study results based on their statistical significance they induce `dissemination bias'. The result is a skewed distribution of findings in the public domain that does not represent the true position in a research field.

My collaborators and I will dedicate our research towards solving these problems. Together we will develop statistical theory, put it to the test using computer simulation, and then present our results at international conferences for further critical appraisal. Once we are confident of their worth, we will apply our new methods in the analysis and interpretation of existing patient data. In the long term our work may influence the way that future scientific studies are designed, reduce wasted resources in the NHS and, ultimately, benefit the health and wellbeing of society.

My methodological research will be split into three areas of estimation. Please see `Case for Support' for further clarification.

In Methodology area 1 (G-estimation) I plan to develop:
(1a) Bias-adjusted Structural Mean Models (SMMs) for Mendelian randomisation (MR) analyses using multiple instrumental variables (IVs) that use genome-wide association study (GWAS) data for IV selection;
(1b) SMMs that can detect and, if necessary, adjust for pleiotropy in MR analyses;
(1c) SMMs that are robust to weak instrument bias in MR analyses;
(1d) Optimal time-dependent weights for the Rank Preserving Structural Failure Time Model (RPSFTM) that provide more precise G-estimates of the causal effect of treatment in clinical trials of late stage cancer patients;
(1e) RPSFTMs that are valid in the presence of treatment effect heterogeneity.
Objectives (1a)-(1e) will be in conjunction with Professor's Nuala Sheehan, George Davey Smith, Stijn Vansteelandt and Dr Ian White.

In Methodology area 2 (Minimum variance unbiased estimation) I plan to:
(2a) Extend MVUE methodology from two-stage to multi-stage Adaptive Seamless Designs (ASDs);
(2b) Adapt and apply new MVUE methodology from (2a) to quantify sensitivity, specificity, ROC curve and risk prediction model parameters in multi-stage trials into diagnostic biomarkers.
Objectives (2a-2b) will be in collaboration with Professor's Ekkehard Glimm and Toby Prevost.

In Methodology area 3 (Shrinkage estimation) I plan to develop:
(3a) Methods of shrinkage estimation for selected treatments in two-stage and multi-stage ASDs;
(3b) A common Empirical Bayesian framework for bias-adjusted meta-analysis in the presence of small study effects.
Objectives (3a-b) will be in conjunction with Professor's Werner Brannath, Sue Todd and Dr Gerta Rucker.

Epidemiologists will be able to use my work on Mendelian Randomisation to uncover new causal disease mechanisms. This will provide solid, reliable evidence for public health interventions to be initiated that have a positive impact on the nation's wellbeing. This could, in the long term, save money for the UK government in reduced contributions to the National Health Service.

My work on adjusting for treatment switching will benefit patients, by providing the tools to allow them to be treated ethically through the course of a late stage cancer trial, but for useful information to be extracted at the trial's conclusion. It will also benefit regulatory bodies such as the National Institute of Clinical Excellence (NICE), by enabling them to judge whether the treatment is cost-effective enough to licence across the NHS. This would most recently have helped NICE to better assess treatments for gastrointestinal stromal tumours and renal cell carcinoma. A better informed NICE can advise the NHS to use its finite budget in a way that best cares for the UK's health needs.

My work on unbiased estimation will enable pharmaceutical companies to conduct more powerful and cost efficient multi-stage adaptive trials in search of effective treatments. The evidence from such trials would not be accepted by regulatory authorities such as the FDA unless these estimates can be reported. By reducing the cost taken to develop new treatments, pharmaceutical companies can charge less for them in order to recuperate their initial investment. This could save the NHS and UK government money on future prescription charges.

By developing methods for meta-analyses that are robust to small study effects, I will enable medical professionals to obtain a balanced picture of the research landscape when considering what is `best care' for patients. If they are seen to be taking a more rigorous, measured approach to evidence synthesis, this could increase the general public's level of trust in the profession as a whole.