Assurance methods for adaptive clinical trial designs

在设计临床试验时，必须仔细计划招募的患者数量。如果数量太少，试验就有可能不能提供足够的证据证明治疗有效。如果数量过多，一些患者将被不必要地纳入研究，可能会接受当时可能被证明无效的治疗。传统上，在选择患者数量时，使用了统计学上的“权力”概念。非正式地说，临床试验的力量是试验成功并证明治疗有效的可能性。关键的是，power假设新的治疗方法确实如预期的那样有效。但在研究开始之前，我们无法知道这是真的。实际上，试验成功的频率比功率计算预测的要低得多，这是非常昂贵的。这个项目将发展另一种统计方法：“保证”，即在进行试验之前对治疗效果的不确定性进行适当评估，以便对试验成功的机会作出更现实的评估。具体地说，将为更复杂类型的试验设计开发保证方法。在保证方法中，首先规定试验的格式和分析试验数据的方法，就像常规的功率计算一样。然后我们从专家那里“引出”一个概率分布——在现有证据和专业知识的支持下——来表示对新疗法有效性的不确定性。有了这个分布，我们就可以计算出试验成功的概率，考虑到我们目前对治疗效果的不确定性。从专家那里得到概率分布的一般技术用于贝叶斯统计（获得“先验”分布）和概率风险分析。专家启发的挑战是如何将特定领域的知识和不确定性转换为概率分布，特别是当我们需要统计模型中某些参数的分布时，专家可能会发现难以直接评估。其他问题包括当不同专家意见不一致时该怎么办，以及如何证明最终选择的概率分布。只有少数关于保证方法的出版物，这些都假设非常简单的试验设计和分析，例如，只有一个治疗组，一个对照组，数据将用单独的双样本t检验进行分析。该项目的目的是开发更复杂的试验设计，如自适应设计的保证方法。学生将首先对临床试验设计和现有的保证方法进行文献综述，并与行业合作伙伴协商，选择一些临床试验设计用于保证方法的开发。将指定在特定试验设计中收集的数据类型，并确定用于分析数据的统计模型。然后识别模型中的不确定参数。对于选定的设计，可交付成果将是1。一个启发方案，列出要向（特定治疗）专家提出的问题，以便可以为模型中所有不确定参数构建概率分布。计算保证的计算方法规范。适应目前的设计方案，利用引出的保证。实现基于web的应用程序的方法，用R包制作的闪闪发光。

When designing a clinical trial, the number of patients recruited to the trial must be planned for carefully. If there are too few, there is a risk that the trial will not provide sufficient evidence that the treatment works. If there are too many, some patients will be needlessly enrolled into a study, potentially receiving a treatment that could, at that time, be demonstrated to be ineffective.Traditionally, to choose the number of patients, the statistical concept of 'power' is used. Informally, the power of a clinical trial is the probability that the trial will be successful and demonstrate that the treatment works. Critically, power assumes the new treatment really does work as well as desired. But we cannot know this to be true before the study starts. In practice, trials are successful significantly less frequently than predicted from power calculations, and this is very costly.This project will develop an alternative statistical method: "assurance" that involves properly assessing uncertainty about the effectiveness of the treatment, before the trial is conducted, so that a more realistic assessment can be made of the chances of success for the trial. Specifically, assurance methods will developed for more complex types of trial design.In the assurance method, the format of the trial and the method for analysing the trial data are first specified, as they would be for a conventional power calculation. We then 'elicit' a probability distribution from experts - supported by available evidence and expertise - to represent uncertainty about the effectiveness of the new treatment. Given this distribution, we can compute the probability the trial will be successful, allowing for our current uncertainty about how well the treatment works. The general technique of eliciting a probability distribution from experts is used in Bayesian statistics (to obtain a 'prior' distribution) and in probabilistic risk analysis. The challenge in expert elicitation is how to convert domain-specific knowledge and uncertainty to a probability distribution, in particular when we need a distribution for some parameter in a statistical model that the experts may find difficult to assess directly. Other problems include what to do when there are different experts who disagree with each other, and how to justify the final choice of probability distribution.There have only been a small number of publications on assurance methods, and these all assume very simple trial designs and analysis, for example, that there will be a single treatment group, a single control group, and that the data will be analyses with a solitary two-sample t-test. The aim of this project is to develop assurance methods for more complex trial designs such as adaptive designs. The student will first conduct a literature review of clinical trial designs and existing assurance methods, and in consultation with the industrial partner, choose some clinical trial designs for the development of assurance methods. The type of data to be collected within a particular trial design will be specified, and the statistical model used to analyse the data will be determined. The uncertain parameters in the model will then be identified. For the chosen design, the deliverables will be1. an elicitation protocol, setting out what questions to ask the (treatment-specific) experts, such that probability distributions can be constructed for all the uncertain parameters in the model.2. specification of computational methods for calculating assurances.3. adaptation of the current design protocol, to exploit the elicited assurance.4. web-based apps to implement the methods, produced with the R package shiny.