Pleiotropy and principal components of heritability combine to increase power for association analysis
Pleiotropy and principal components of heritability combine to increase power for association analysis
复制标题
DOI：
10.1002/gepi.20257
复制
发表时间：
2008-01-01
影响因子：
2.1
通讯作者：
Roeder, Kathryn
中科院分区：
文献类型：
--
作者：
Klei, Lambertus;Luca, Diana;Roeder, Kathryn
当测量许多相关性状时，存在通过基因型多态性发现这些性状的协调控制的潜力。一个常见的统计方法来解决这个问题，包括评估每个表型和每个单核苷酸多态性（SNP）单独（PHN）之间的关系;并采取邦弗罗尼校正进行独立测试的有效数量。或者，可以应用降维技术，例如主成分估计，并测试与表型的主成分（PCP）而不是个体表型的关联。在Lange及其同事的工作基础上，我们开发了一种基于遗传力主成分（PCH）的替代方法。对于每个SNP，PCH方法将表型减少到具有比表型的任何其他线性组合更高的遗传力的单个性状。因此，SNP与衍生性状之间的关联通常比与任何个体表型或PCP之间的关联更容易检测。当应用于不相关的主题时，PCH有一个缺点。对于每个SNP，有必要估计使所有表型的遗传力最大化的负载向量。我们开发了一种迭代样本分割的方法，使用一部分数据进行训练，其余部分用于测试。这种交叉验证方法保持了I型错误控制，但有效地利用了数据，从而产生了强大的关联测试。
When many correlated traits are measured the potential exists to discover the coordinated control of these traits via genotyped polymorphisms. A common statistical approach to this problem involves assessing the relationship between each phenotype and each single nucleotide polymorphism (SNP) individually (PHN); and taking a Bonferroni correction for the effective number of independent tests conducted. Alternatively, one can apply a dimension reduction technique, such as estimation of principal components, and test for an association with the principal components of the phenotypes (PCP) rather than the individual phenotypes. Building on the work of Lange and colleagues we develop an alternative method based on the principal component of heritability (PCH). For each SNP the PCH approach reduces the phenotypes to a single trait that has a higher heritability than any other linear combination of the phenotypes. As a result, the association between a SNP and derived trait is often easier to detect than an association with any of the individual phenotypes or the PCP. When applied to unrelated subjects, PCH has a drawback. For each SNP it is necessary to estimate the vector of loadings that maximize the heritability over all phenotypes. We develop a method of iterated sample splitting that uses one portion of the data for training and the remainder for testing. This cross-validation approach maintains the type I error control and yet utilizes the data efficiently, resulting in a powerful test for association.