Collaborative Research: Modernizing Mixed Model Prediction

从医学到经济，从商业到社交媒体，社会许多领域的信息爆炸给现代数据科学带来了关于学科层面知识的紧迫问题，例如精准医学、集中营销、家庭经济学和许多其他领域。其中包括在重要应用领域进行数据分析和预测的有效方法，从通过差分隐私（DP）来保护隐私，到专注于预测表观遗传标记的精准医疗和公共卫生差异，以及利用美国劳工统计局（BLS）的就业数据进行预测。该项目旨在开发和采用被称为混合模型预测的新方法。特别是，对于DP申请，调查人员将把方法应用于公开发布的2020年美国十年一次的人口普查；对于劳工统计局的应用，研究人员将针对正在进行的COVID-19大流行期间的波动性问题，因此需要对传统方法进行强有力的修改。该研究将与沉浸在特定应用领域的合作者一起进行。在该项目中，研究人员将关注三个主要目标：1)基因组预测问题中的多变量混合模型预测（MMP），其中相关DNA甲基化标记反映了潜在的疾病生物学，并且通过借鉴这种多变量结构的优势可以提高预测准确性；2)差分私有（DP）数据的MMP，其中集群或组身份被设计污染，不被释放以保护隐私；3)具有非高斯随机效应和误差的MMP，这极大地扩展了MMP可以应用的情况范围，超出了经典正态性假设，这些假设不适合许多现代数据集。研究人员将为每个目标开发所需的方法，从理论上研究程序，并进行广泛的经验模拟研究，将新方法与其他方法进行比较。此外，研究人员将与他们在主题领域的合作者密切合作，实施本项目开发的方法来回答实际问题。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。

The information explosion in many areas of society, from medicine to economics and business to social media, has resulted in pressing questions for modern data science regarding subject-level knowledge, such as in precision medicine, focused marketing, family economics, and many other areas. These include effective methods for data analysis and prediction in important areas of application ranging from privacy protection via differential privacy (DP) to precision medicine and public health disparities focusing on the prediction of epigenetic markers, and to predictions with employment data from the U.S. Bureau of Labor Statistics (BLS). This project aims to develop and employ new methods known as mixed model prediction. Particularly, for the DP application, the investigators will apply the methods to the publicly released 2020 U.S. decennial census; for the BLS application the investigators will target questions regarding volatility during the ongoing COVID-19 pandemic that thus require robust modifications from traditional approaches. The research will be carried out in conjunction with collaborators who are immersed in a particular application area.In this project, the investigators will focus on three major aims: 1) multivariate mixed model prediction (MMP) in genomic prediction problems where correlated DNA methylation markers reflect underlying disease biology and improved prediction accuracy is possible by borrowing strength across this multivariate structure; 2) MMP for differentially private (DP) data in which cluster or grouping identities are contaminated by design and not released to protect privacy; and 3) MMP with non-Gaussian random effects and errors, which greatly can expand the range of circumstances in which MMP can be applied beyond the classical normality assumptions that do not fit many modern datasets. The investigators will develop the required methodology for each aim, study the procedures theoretically, and carry out extensive empirical simulation studies to compare the new methods with other methods. Furthermore, the investigators will work closely with their collaborators in the subject fields on implementing the methods developed in this project to answering practical questions.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.