HOD1: Inferring relative treatment effects from combined randomised and observational data

点击翻译按钮获取中文摘要

Randomised controlled trial evidence is the recognized "gold-standard" to estimate relative treatment effects to inform decision making, because patients are allocated to interventions randomly which guarantees that any patient characteristics that may be related to the patients' outcome are not related to the intervention they were allocated (since this was randomly chosen). Thus, patients randomised to different groups differ only in their allocated intervention and any difference in outcomes can be attributed to the intervention alone.In the absence of randomisation, treatment assignment may be based on patient characteristics that also lead to a different outcome. This makes it unclear whether the different outcomes were due to the different interventions or to different patient characteristics.Methods for combining the results of many randomised trials and obtaining a combined measure of the effect of the interventions of interest are well established. However, when randomised evidence is limited in quantity or quality, as is increasingly the case, it is natural to consider the use of non-randomised (observational) data as a substitute or supplementary source of evidence. The inclusion of observational evidence raises challenges as this evidence may not be estimating the true effect of the intervention, it may be biased.Different methods have been proposed to handle this bias according to the origin and structure of the evidence available. However, so far, no clear consensus has emerged regarding any of the methods proposed. Thus, Health Technology Assessment bodies and particularly reimbursement agencies such as NICE face increasing challenges in their assessments of evidence and have called for extra research. To answer this call, we propose to carry out five interlinked projects: four of them are focused each on a different type of evidence structure. A fifth project develops methods that examine the impact of bias, or potential bias, on the treatment decision.The first project will evaluate the degree of error in the treatment effects produced when including studies with a single intervention (non-comparative studies), or where the interventions of interest can be separated into two groups of interventions where no study has compared an intervention in one group with an intervention in the other group. This project relies on there being individual participant data from at least one of the studies. The second project will consider cases where the individual participant data are not available and will (1) investigate the properties of the various methods; (2) determine which of the methods is likely to be best at reducing decision uncertainty for different kinds of observational data.The third project will carry out computer simulations where the 'true' treatment effects can be assumed to be known and where the simulated (computer generated) data are realistically representative of the real data of interest to assess the circumstances under which the different methods are most likely to produce the least biased results. The fourth project is specifically oriented to assess managed entry of new pharmaceuticals. There is a trend towards appraising new, promising, health technologies early, when available evidence is still limited. Methods used to inform these decisions make strong assumptions. We will explore the potential for observational data available in routinely collected databases to validate these assumptions.The final project will extend existing methods to answer the question: "suppose this evidence is biased, how biased would it have to be before it changed our decision as to which is the best treatment?" where the decision can be based on, for example, the treatment which represents best value for money to the NHS.