Collaborative Research: Subject-level Prediction and Application

许多实际问题与预测有关，其中主要兴趣是在主题（例如个性化或精确医疗）或（小）亚人群（例如小社区）水平上。近年来，商业、社会科学和健康科学等不同领域出现了新的和具有挑战性的问题。例如，可以预测新患者的健康结果，或者预测新学校对儿童预防吸烟教育的反应。研究人员在之前的工作中表明，分类混合模型预测（CMMP），在这种情况下，通过确定新受试者所属的类别，有可能在预测准确性方面取得实质性进展。然而，CMMP目前运行的场景在一定程度上受到限制，许多现实生活中的情况都超出了它的范围。考虑到在准确性方面可能获得的巨大收益，开发进一步的方法和计算进步来深化这一领域的知识将是非常有价值的。本项目旨在将分类混合模型预测方法的方法论进展推广到其他类型的学科级预测问题中，并沿着CMMP思想开发新的推理方法，使后者在实际情况中真正有用。CMMP的基本思想是在希望进行预测的总体中的组或集群与具有已知组或集群的（大规模）训练数据集之间创建“匹配”。一旦建立了这样的匹配，就可以利用传统的混合模型预测方法进行准确的预测。本课题要解决的实际挑战包括：1)如何处理具有未知分组的训练数据；Ii)如何处理稀疏的高维协变量；iii)如何更好地利用协变量信息来提高CMMP的准确性；iv)如何为cmmp型预测提供准确的不确定性测量。我们将研究两个重要的应用领域。一个是精准医学和健康差异，重点是使用高维基因型谱预测表观遗传标记。另一个来自家庭经济学领域，使用了来自中国的大量数据调查，其中对更精细的分辨率水平（例如，家庭）的预测是主要兴趣。这两个应用程序都将利用与从业者的重要合作，从而增加工作的影响。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

Many practical problems are related to prediction, where the main interest is at the subject (for example personalized or precision medicine) or (small) sub-population (for example small community) level. In recent years, new and challenging problems have emerged from diverse fields such as business, social sciences, and health sciences. Examples may involve prediction of a health outcome for a new patient or perhaps prediction of a new school's response to efforts to educate children about smoking prevention. The investigators have shown in previous work called classified mixed model prediction (CMMP) that in such cases, it is possible to make substantial gains in prediction accuracy by identifying a class that a new subject belongs to. However, the scenarios under which CMMP currently operates are somewhat constrained and many real-life situations fall outside its scope. Given the tremendous gains in accuracy that are possible, it would be very valuable to develop further methodology and computational advances to deepen knowledge in this area. This project aims to make methodological advances of the classified mixed model prediction method into other types of subject-level prediction problems as well as to develop new inferential methods along the CMMP idea, by making the latter truly useful in practical situations. The basic idea of CMMP is to create a "match" between a group or cluster in the population for which one wishes to make prediction and a (massive) training dataset, with known groups or clusters. Once such a match is built, the traditional mixed model prediction method can be utilized to make accurate predictions. The practical challenges that will be solved in this project include i) how to deal with training data with unknown grouping; ii) how to deal with sparse, high dimensional covariates; iii) how to make better use of covariate information to improve accuracy of CMMP; and iv) how to provide accurate measures of uncertainty for CMMP-type predictions. Two important areas of application will be investigated. One is in precision medicine and health disparities focusing on the prediction of epigenetic markers using high dimensional genotype profiles. The other comes from the area of family economics using a large survey of data from China where predictions at finer levels of resolution (e.g., households) are of primary interest. Both applications will leverage important collaborations with practitioners and thus increase the impact of the work.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.