Effective population adjustment in evidence synthesis of randomised controlled trials for health technology assessment

为了决定向患者推荐哪些治疗方法，我们需要对不同治疗方法之间的比较进行可靠的估计。然而，直接比较所有感兴趣的治疗方法的研究可能并不存在。取而代之的是，我们经常有一系列研究比较不同的治疗方法，或者在某些情况下只比较一种治疗方法。此外，在改变治疗效果的不同研究中，患者之间可能存在差异。为了解决这些问题，出现了一种称为“多水平网络元回归”(ML-核磁共振)的统计方法。这种方法结合了来自多项研究的证据，其中一些研究提供了每个参与者的个人水平数据，而另一些研究只提供了已公布的概要估计，并考虑了患者群体之间的差异--这一过程被称为“人口调整”。重要的是，这种方法可以产生特定于相关人群的估计以供决策(例如，英国患者群体)。这意味着，国家健康与护理卓越研究所(NICE)等决策者可以针对相关人群做出更好的决策。然而，如果要更广泛和有效地将ML-核磁共振用于决策，在实践中使用ML-核磁共振有几个障碍需要解决。首先，该方法需要关于每种治疗的大量数据，而这些数据并不总是可用的。例如，一家向NICE提交申请的公司很可能拥有自己治疗试验的个人水平数据，但只公布了竞争对手的试验总结。如果没有足够的数据，我们可能会试图通过对不同的治疗组如何工作做出假设来简化统计模型，但这些假设可能不合适，这可能会导致结果中的系统性错误和得出错误的结论。其次，临床试验经常会遇到一些问题，如数据缺失、参与者没有接受分配给他们的治疗，或者参与者被允许切换治疗(例如，如果他们的疾病进展)。统计方法可以用来解释这些问题，因为如果处理不当，可能会导致结果中的系统性误差。然而，目前这些方法不能与ML-核磁共振等方法一起使用来解释种群之间的差异。该项目旨在解决这些问题，以确保ML-核磁共振在决策者最经常遇到的情况下工作良好。这将通过以下方式实现：i)为ML-核磁共振开发新的统计方法，以使用从已发表的试验报告中获得的附加信息；ii)提出更新临床试验如何报告的指南的建议，以改进已发表报告中这些附加信息的可用性；iii)通过真实和模拟的例子来调查统计方法的性能；iv)开发新的统计方法，将总体调整与解决临床试验中常见问题的方法相结合，例如丢失数据或切换治疗；以及v)开发可获得的软件工具和培训课程，以支持采用这些方法。这项研究将对NICE等决策者产生直接影响，并将导致更好地做出知情的治疗决定。统计方法的拟议进步和报告临床试验的更新建议有可能在更广泛的背景下改变医疗决策，即使只有已公布的摘要数据可用，例如制定NICE临床指南。此外，在个性化医疗方面也有直接的应用，推荐对象是个人或较小的群体。

To decide which treatments to recommend to patients, we need reliable estimates of how the different treatments compare to each other. However, studies that directly compare all treatments of interest may not be available. Instead, we often have a mixture of studies that compare a selection of different treatments, or in some cases only a single treatment. Furthermore, there may be differences between the patients in the different studies that change how well the treatments work. To address these issues, a statistical method called "multilevel network meta-regression" (ML-NMR) is available. This method combines evidence from multiple studies, where some studies provide individual-level data on every participant and some only provide published summary estimates, and accounts for differences between patient populations - a process known as "population adjustment". Importantly, this method can produce estimates that are specific to a relevant population for decision-making (e.g. the UK patient population). This means that decision makers such as the National Institute for Health and Care Excellence (NICE) can make better decisions that are targeted to the relevant population. However, there are several barriers to the use of ML-NMR in practice which need to be addressed if it is to be used more widely and effectively for decision-making. Firstly, the method requires substantial amounts of data on each treatment, which are not always available. For example, a company making a submission to NICE is likely to have individual-level data from their own trials of their own treatment, but only published summaries from their competitors' trials. Without enough data, we may instead attempt to simplify the statistical model by making assumptions about how different groups of treatments work, but these assumptions may not be appropriate, which can lead to systematic errors in the results and the wrong conclusions being drawn. Secondly, it is common for clinical trials to encounter issues such as missing data, participants not receiving the treatment they were assigned, or participants being allowed to switch treatments (e.g. if their disease progresses). Statistical methods are available to account for these issues, since if they are not handled correctly they can lead to systematic errors in the results. However, currently these methods cannot be used together with methods to account for differences between populations like ML-NMR. This project aims to address these issues to ensure that ML-NMR works well in situations most frequently encountered by decision makers. This will be achieved by: i) developing novel statistical methods for ML-NMR to use additional information available from published trial reports; ii) making recommendations to update guidelines for how clinical trials are reported, to improve the availability of this additional information in published reports; iii) investigating the performance of the statistical methods through real and simulated examples; iv) developing novel statistical methods to combine population adjustment with methods that account for common issues in clinical trials such as missing data or switching treatments; and v) developing accessible software tools and training courses to support the uptake of the methods.This research will have direct impact for decision makers such as NICE and will lead to better informed treatment decisions. The proposed advances in statistical methods and updated recommendations for reporting clinical trials have the potential to transform healthcare decision-making in wider contexts, even when only published summary data are available, such as the development of NICE clinical guidelines. Additionally, there are direct applications in personalised medicine, where recommendations are targeted to individuals or smaller groups.