Collaborative Research: Prediction and Model Selection for New Challenging Problems with Complex Data+

协作研究：复杂数据新挑战性问题的预测和模型选择

• 批准号: 1510219
• 负责人: Jiming Jiang
• 金额: $ 11.23万
• 依托单位: University of California-Davis
• 依托单位国家: 美国
• 项目类别: Standard Grant
• 财政年份: 2015
• 资助国家: 美国
• 起止时间: 2015-08-15 至 2018-07-31
• 项目状态: 已结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=1510219&HistoricalAwards=false
• 关键词: Collaborative; Research; Prediction; Model; Selection

Mixed model prediction, that is, prediction based on a class of statistical models known as mixed effects models, has a fairly long history. The traditional fields of applications have included genetics, agriculture, education, and surveys. Nowadays, new and challenging problems have emerged from such fields as business and health sciences, in addition to the traditional fields, to which methods of mixed model prediction are potentially applicable, but not without further methodology and computational developments. Some of these problems occur when interest is at subject level, such as personalized medicine, or (small) sub-population level, such as small communities, rather than at large population level. In such cases, it is possible to make substantial gains in prediction accuracy by identifying a class that a new subject belongs to. Other challenging problems occur when applying existing model search strategies in situations of incomplete or missing data, in model search or selection when prediction is of primary interest, and in making statistical inference based on the result of model search or selection. This collaborative research project aims at solving these challenging problems in prediction and model selection in situations of complex data, such as incomplete or missing data, and data that are correlated due to presence of random effects.In this collaborative research project the PIs develop a novel statistical method, called classified mixed model prediction, to identify the subject class. This way, the new subject is associated with a random effect corresponding to the same class in the training data, so that the mixed model prediction method can be used to make the best prediction. Furthermore, the PIs develop a recently proposed method, called E-MS algorithm, for model selection in the presence of incomplete or missing data. The PIs also develop an idea called predictive model selection by deriving a predictive measure of lack-of-fit, and combining this measure with a recently developed class of strategies of model selection, called the fence methods. Finally, the PIs develop a unified Jackknife method to accurately assess uncertainty in mixed model analysis after model selection. Theories will be established for these new methods, and their performance and potential gains through extensive Monte-Carlo simulations will be studied. The new methods will be implemented in the R language/environment for statistical computing and graphics. All of the developed methodologies will be applied and tested in a number of applications via a series of close collaborations with experts who will provide access to the data and also guidance in interpretation and dissemination of findings. The fields of applications include genetics, health and medicine, agriculture, education, business and economy. The research project will also promote teaching, training and learning that involve under-represented groups, and build research networks between our institutions.

混合模型预测，即基于一类被称为混合效应模型的统计模型的预测，具有相当长的历史。传统的应用领域包括遗传学、农业、教育和调查。如今，新的和具有挑战性的问题已经出现，从商业和健康科学等领域，除了传统的领域，混合模型预测的方法是潜在的适用，但不是没有进一步的方法和计算的发展。其中一些问题发生在兴趣处于学科水平（如个性化医疗）或（小）亚人群水平（如小社区）而不是大人群水平时。在这种情况下，可以通过识别新对象所属的类别来大幅提高预测精度。当在不完整或缺失数据的情况下应用现有模型搜索策略时，在预测是主要兴趣的模型搜索或选择中，以及在基于模型搜索或选择的结果进行统计推断时，会出现其他具有挑战性的问题。本合作研究项目旨在解决在数据不完整或缺失、随机效应相关等复杂数据情况下的预测和模型选择等难题。在本合作研究项目中，PI开发了一种新的统计方法，称为分类混合模型预测，用于识别主题类别。这样，新的主题与训练数据中对应于同一类的随机效应相关联，从而可以使用混合模型预测方法进行最佳预测。此外，PI开发了一种最近提出的方法，称为E-MS算法，用于在存在不完整或缺失数据的情况下进行模型选择。PI还开发了一种称为预测模型选择的想法，通过推导出一种预测性的失拟度量，并将这种度量与最近开发的一类模型选择策略（称为栅栏方法）相结合。最后，PI开发了一种统一的Jackknife方法，以准确评估模型选择后混合模型分析中的不确定性。将为这些新方法建立理论，并通过广泛的蒙特-卡罗模拟研究其性能和潜在收益。新方法将在R语言/环境中实现，用于统计计算和图形。将通过与专家的一系列密切合作，在若干应用程序中应用和测试所有开发的方法，这些专家将提供数据访问以及解释和传播调查结果的指导。应用领域包括遗传学、健康和医学、农业、教育、商业和经济。该研究项目还将促进涉及代表性不足群体的教学，培训和学习，并在我们的机构之间建立研究网络。