Collaborative Research: Detecting false discoveries under dependence using mixtures

多个测试问题的统计分析围绕着同时测试产生的p值集合的分布。来自功能磁共振成像、蛋白质组学、微阵列和其他生物医学实验的数据显示了p值之间的相关性。如果在估计诸如错误发现率之类的差错控制措施时不考虑这种相关性，则统计推断会产生生物学上不相关的结论。这一建议描述了一种面向模型的方法，通过灵活和准确地利用混合对相关情况下的p值的联合分布进行建模来检验多个假设。研究人员的另一个理论目标是研究倾斜混合模型的性质。通过将相关性纳入p值模型，建议的研究提供了对错误发现的有效控制，特别是在复杂的生物医学应用中。多假设检验是科学家有效分析大规模复杂生物医学数据(如基因表达数据、蛋白质数据或脑成像数据)的主要统计工具之一。在这样的生物医学应用中，疾病关联研究需要同时测试数千个基因或蛋白质或大脑区域的关联意义。识别一种基因或一种蛋白质可能与某一特定疾病有关的现象称为发现。然而，在大规模的生物医学研究中，存在着通过做出太多错误发现而积累错误的风险。这项拟议的研究通过提供对大规模疾病关联研究中错误率的准确估计，极大地影响了生物医学应用中的统计学实践。研究人员专门为自闭症患者的MRI研究中的脑成像应用开发了错误控制机制。该项目以研究生教育和培训的形式影响人力资源开发。

Statistical analysis of multiple testing problems revolves around the distribution of the collection of p-values arising from simultaneous tests. Data from fMRI, Proteomics, Microarray and other biomedical experiments exhibit dependence among p-values. Statistical inference yields biologically irrelevant conclusions if such dependence is not taken into consideration while estimating error control measures such as the false discovery rate. This proposal delineates a model oriented approach to multiple hypotheses testing by flexible and accurate modeling of the joint distribution of the p-values in dependent situations using mixtures. An additional theoretical goal the investigators study properties of skew-mixture models. By incorporating dependence in the model for the p-values, the proposed research provides valid controls of false discoveries, especially in complex biomedical applications. The proposed methodologies provide a foundation for statistical analysis in large dependent multiple testing situations and will spawn new research in the area of false discovery control.Multiple hypothesis testing is one of the primary statistical tools available to the scientists for efficiently analyzing large-scale complex biomedical data such as gene-expression data, protemics data or brain imaging data. Disease association studies in such biomedical applications require testing significance of association of several thousand genes or proteins or brain regions, simultaneously. Identification of a gene or a protein as being potentially associated with a given disease is called a discovery. However, in large scale biomedical studies there is a risk of accumulating error via making too many false discoveries. The proposed research substantially influences the practice of statistics in biomedical applications by providing accurate estimates of error rates in large scale disease association studies. The investigators specifically develop error control mechanism for brain imaging applications in MRI studies of autistic patients. The project impacts human resource development in the form of graduate student education and training.