A Robust and Efficient Statistical Framework for Handling Missing-Not-At-Random Data in Patient Reported Outcomes and Beyond

患者报告结果（PRO）代表患者的健康状况，直接来自患者，无需临床医生或其他任何人解释，具有从患者角度描述健康状况的独特特征；因此，PRO 研究对于明智的临床和政策决策以及提高医疗保健的质量和效率具有巨大的希望。然而，PRO 的质量和价值取决于许多因素，其中之一就是非随机缺失 (MNAR) 问题。例如，患者可能因其抑郁程度而无法填写抑郁症调查，或者病情较重的患者可能不太可能完成生活质量调查问卷。一般来说，这些PRO是由于患者健康状况下降而缺失的，但下降的程度是未知的，因为没有观察到；因此，这些缺失的数据具有丰富的信息并且是 MNAR。类似的情况也出现在大规模的健康调查和电子健康记录数据库中。 在这个项目中，PI 将研究 PRO 以及其他类似情况下 MNAR 问题的统计方法和计算算法。该研究产品有潜力应用于各种研究，如阿尔茨海默病、精神健康障碍、骨科和疼痛研究。 PI还将从事学科和跨学科层面的教育，受益者包括当地高中生和本科生、硕士生和博士生以及生物医学研究人员。该项目还将为博士后学者提供研究机会。该项目的总体目标是建立一个突破性的转化统计方法框架，包括稳健的方法和高效的估计器，其中对缺失数据机制的假设被施加在最低水平，因此所开发的方法可以以最大的灵活性应用。由于没有足够的方法来测试缺失数据机制的正确性这一众所周知的事实，PI将采用影子变量方法来实现模型识别，并且本质上不对机制做出进一步的假设，从而为模型错误指定提供最大可能的保护。该方法通过利用基于模型的可能性及其相关的半参数结构，对机制模型错误指定具有鲁棒性。该项目开发的统计方法将被实施到高效的 R 包和用户友好的界面中，供主要目标是分析缺失数据（尤其是 MNAR 数据）的研究人员使用。该奖项反映了 NSF 的法定使命，并通过使用基金会的智力价值和更广泛的影响审查标准进行评估，被认为值得支持。

The patient-reported outcome (PRO), representing the status of the patient's health that comes directly from the patient without interpretation by the clinician or anyone else, has the unique feature of describing health status from the viewpoint of the patient; therefore, the PRO research holds great promise for informed clinical and policy decision-making, as well as for improving the quality and efficiency of healthcare. However, the quality and value of PRO is contingent on a number of factors, and one of them is the missing-not-at-random (MNAR) issue. For instance, patients might fail to fill in a depression survey because of their level of depression, or patients who are sicker may be less likely to complete a quality-of-life questionnaire. In general, these PROs are missing due to the patient's declining health status, but the extent of decline is not known because it is not observed; hence, these missing data are informative and are MNAR. Similar situations also appear in large-scale health surveys and electronic health records database. In this project, the PIs will study statistical methodology and computational algorithm for the MNAR issue in PRO as well as in other similar situations. The research product has the potential to be applied to various studies, such as Alzheimer's disease, mental health disorders, orthopedics, and pain research. The PIs will also engage in education at both disciplinary and interdisciplinary levels, with beneficiaries ranging from local high school students and undergraduates, to master and PhD students, and to biomedical investigators. The project will also provide research opportunities for postdoctoral scholars. The overarching goal of this project is to establish a groundbreaking and translational statistical methodology framework including robust methods as well as efficient estimators, where the assumption on the missing data mechanism is imposed at a minimum level hence the developed methods can be applied with the largest flexibility. Motivated by the well-recognized fact that there is no adequate way to test the correctness of the missing data mechanism, the PIs will adopt the shadow variable approach to achieve the model identification and essentially make no further assumptions on the mechanism, thereby provide largest possible protection to model misspecification. The methodology is robust against the mechanism model misspecification by leveraging the model-based likelihood and its associated semiparametric structure. The statistical methods developed in this project will be implemented into efficient R packages and user-friendly interfaces for researchers whose primary goal is the analysis of missing data, especially MNAR data.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.