Statistical Methods for Evaluating Markers for Treatment Selection

描述（由申请人提供）：治疗选择标志物有可能改善患者结局并降低医疗成本。当一种治疗方法只对一部分患者有益时，可以使用一种标记来识别这些受试者，以避免其他人接受不必要的治疗。如果一种疗法对某些人特别有害，一种有用的标记物将识别这些受试者，以避免治疗他们。新技术正在产生大量的候选标记。然而，他们的评估标准，这是必不可少的决策标记的进步和监管批准，是非常缺乏的。本申请旨在为这些标准的制定做出贡献。要求1（“性能测量”）证明了当前评估治疗选择标记物的方法的不足，并开发了标记物性能的三种新的统计测量：（i）标记物-治疗预测性曲线显示了每个标记物值的治疗效果;（ii）治疗选择ROC曲线显示了标记物区分从治疗中受益和不受益的个体的准确性;和iii）选择影响曲线描述了使用标记选择治疗的群体影响。目标2（“比较标记物”）建立在这种方法的基础上，以开发用于比较两个候选标记物的性能的方法。在固定和优化的标记阈值，以及标记性能的全球摘要的比较，提出。目标3（“协变量特异性性能”）开发了一种方法，用于评估标记物性能如何随患者特征或标记物测量程序的各个方面等因素而变化。因为通常寻求标记组合以期望提高性能，所以目标4（“组合标记”）开发了一种组合多个标记并评估组合的性能的方法。这也导致了一种用于评估通过将新标记物添加到现有标记物或临床信息中而获得的性能增量的方法。目标5（“研究设计”）考虑了这些新方法对研究设计的影响。虽然理想的设计是盲法和随机试验，其中在基线时测量所有参与者的标志物，但仔细选择测量标志物的试验受试者子集（例如巢式病例对照设计）可能会产生类似的效率。将制定这两种类型研究的设计方法，包括功效计算和关于匹配和分层的建议。还将提供用于评价在试验参与者子集上测量标志物的设计中的标志物的方法。这项研究将与标志物评估和临床试验设计和分析领域的国际领导者合作进行。这些方法将应用于几个重要的干预试验，其中标记物已被测量用于治疗选择。

DESCRIPTION (provided by applicant): Markers for treatment selection have the potential to improve patient outcomes and decrease medical costs. When a treatment benefits only a subset of patients, a marker that identifies these subjects could be used to spare others unnecessary treatment. If a therapy is particularly harmful to certain individuals, a useful marker would identify these subjects to avoid treating them. New technologies are producing an abundance of candidate markers. However, the standards for their evaluation, which are essential for making decisions regarding marker advancement and regulatory approval, are sorely lacking. This application proposes to contribute to the development of these standards. Aim 1 ("Measures of Performance") demonstrates the inadequacy of the current approach to evaluating treatment selection markers, and develops three novel statistical measures of marker performance: (i) Marker-by-treatment predictiveness curves display the treatment effect at each marker value; ii) Treatment selection ROC curves show the accuracy with which the marker discriminates between individuals who do and do not benefit from treatment; and iii) The selection impact curve describes the population impact of using the marker to select treatment. Aim 2 ("Comparing Markers") builds on this approach to develop methods for comparing the performance of two candidate markers. Comparisons at fixed and optimized marker thresholds, as well as global summaries of marker performance, are proposed. Aim 3 ("Covariate-Specific Performance") develops an approach to evaluating how marker performance varies with factors such as patient characteristics or aspects of the marker measurement procedure. Because marker combinations are commonly sought with the hopes of improving performance, Aim 4 ("Combining Markers") develops an approach to combining multiple markers and evaluating the performance of the combination. This also leads to a method for assessing the increment in performance gained by adding a new marker to existing markers or clinical information. Aim 5 ("Study Design") considers the implications of these new methods for study design. While the ideal design is a blinded and randomized trial where the marker is measured at baseline on all participants, careful selection of a subset of trial subjects in which to measure the marker (eg a nested case-control design) may yield similar efficiency. Approaches to the design of both of these types of studies, including power calculations and recommendations regarding matching and stratification, will be developed. Methods for evaluating markers in designs that measure the marker on a subset of trial participants will also be provided. This research will be conducted in collaboration with international leaders in the fields of marker evaluation and clinical trial design and analysis. The methods will be applied to several important intervention trials where markers have been measured for treatment selection.