Adaptation of New Statistical Ideas for Medicine

我们的优秀奖工作将继续包括两个主要部分：参与特定的生物医学研究项目，如NHBLI的FEHGAS研究，以及开发适合于分析大型、复杂数据集的新统计方法。这些努力是相辅相成的，具体项目表明最需要哪些统计方法，也是新方法的测试案例。 例如，FEHGAS研究试图根据SNP数据(以及年龄和性别等背景变量)预测高血压的发病年龄。有550,000个SNP可用于预测，其中大多数将被证明是无用的，这使得这个问题比在表达微阵列的情况下具有更大的挑战性。Efron计划将经验贝叶斯方法从他最近的论文扩展到这一背景下，希望克服由于单个SNP预测能力通常较弱而造成的困难。Olshen计划将CART(计算机辅助回归树)和Bootstrap方法扩展到选择一组有前途的预测性SNP。 大规模的显著性检验，例如在微阵列癌症研究中选择“显著”的基因，已经成为一个密集的统计学发展领域。然而，适当实施的关键问题在文献中仍然含糊不清：适当零假设的选择；比较集的选择(所有550,000个SNP是应该一起进行测试，还是通过染色体单独测试？)；以及相关性的影响。正如进度报告所述，我们在回答这些问题方面取得了一些进展。我们的持续努力是方法论实施和理论发展的结合。 相关性可能会对标准统计技术产生特别严重的影响。一组微阵列是相互独立的吗？结果表明，由于严重的基因相关性，一项涉及20,000个基因的研究的有效样本量减少到了大约17个。我们目前正在开发诊断方法，以发现海量数据集中的关联困难，并评估它们对假设检验、估计和预测的影响。一项20,000个基因微阵列的研究产生了200,000,000个相关性，对于实际的洞察来说，这听起来太大了。但我们在经验的Ba5‘s近似方面正在取得进展，该近似将相关性和影响总结在一个单独的数字中，适合于简单的分析。 20世纪的生物统计学应用在本质上是压倒性的。纯：然而，频率法对于分析由现代生物医学设备产生的大型、复杂的数据集来说变得不切实际，在这些数据集中，数千个参数和数百万个数据点的关系必须被一起考虑。我们正在继续开发经验贝叶斯方法，允许贝叶斯思想在多重推理问题上发挥作用，而不需要科学家提供特定的事先分布。 一个长期项目是了解经验贝叶斯信息在医学研究中积累的速度有多快。错误发现率是在给定观察数据的情况下，对基因(或SNP或元音)为空的贝叶斯后验概率的估计。我们需要观察多少受试者和多少基因才能得到后验概率的准确经验贝叶斯估计？ 根据我们自己的摩尔定律，自20世纪90年代以来，生物医学数据集的大小每隔几年就增加一个数量级。新兴技术(瓷砖阵列、珠子阵列、适体芯片、甲基化阵列、外显子芯片和各种新的成像设备)有望进一步增加，这对计算设备和统计方法都是不利的。我们的长期目标是提供适合海量数据生物医学需求的算法和理论。

Our MERIT award work will continue to have two main components: involvement in .specific biomedical research projects such as NHBLI's FEHGAS study, and development of new statistical methods appropriate for the analysis of large, complex data sets. These efforts are complementary, with the specific projects ¿suggesting which statistical methods are most needed, and also serving as test cases for new methodology. The FEHGAS study, for example,-seeks to predict age of onset of hypertension from SNP data (and background variables such as age and gender). There are 550,000 SNPs available for prediction, most of which will turn out to be useless, making the problem an order of magnitude more challenging, than in expression microarray situations. Efron plans to extend the empirical Bayes methodology from his recent paper to this context, hopefully overcoming the difficulties caused by the usually weak predictive power of individual SNPs. Olshen plans to extend CART (Computer Assisted Regression Trees) and bootstrap methodology to the selection of groups of promising predictive SNPs. Large-scale significance testing, for instance selecting 'significant' genes in a microarray cancer study, has become an area of intense statistical development. Nevertheless, crucial questions of appropriate implementation remain vague in the literature: the choice of an appropriate null hypothesis; the selection of a comparison set (Should all 550,000 SNPs be tested together or separately by chromosome?); and the effects of correlation. We have made some headway in answering these questions, as described in the Progress Report. Our continuing efforts are a combination of methodological implementation and theoretical development. Correlation can have particularly drastic effects on standard statistical techniques. Iii "Are a .set of microarrays independent of each other?" it is shown that a study involving 20,000 genes has its effective sample size reduced to about 17 because of severe gene-wise correlation. We are currently developing diagnostic methods to spot correlation difficulties in massive data sets, and to assess their effects on hypothesis tests, estimates, and predictions. A 20,000 gene microarray study produces 200,000,000 correlations, which sounds oppressively large for practical insight. But we are making progress on an empirical Ba5'es approximation that summarizes correlation, effects in a single number, suitable for simple analysis. Twentieth Century biostatistical applications were overwhelminglyrequentist in nature. Pure: frequentism, though, become impractical for analyzing the large, complex data sets produced by modem biomedical devices, where the relationships of thousands of parameters and millions of data points have to be considered together. We are continuing to develop empirical Bayes methods that allow Bayesian ideas to be brought to bear on questions of multiple inference, without requiring specific prior distributions from the .scientist. A long-term project is to understand how quickly empirical Bayes information accrues in a medical study. A False Discovery Rate is an estimate of the Bayes posterior probability that a gene (or a SNP, or a vowel) is 'null', given the observed data. How many subjects and how many genes do we need to observe in order to get an accurate empirical Bayes estimate of the posterior probability? In our own version of Moore's law, biomedical data sets have increased an order of magnitude in size every few years since the 1990s. Emerging technologies (tiling arrays, bead arrays, aptamer chips, methylation arrays, exon chips, and a variety of new imaging devices) promise further increases, taxing both computational equipment and statistical methodology. Our long-term MERIT goal is to provide algorithms and theory appropriate to massive-data biomedical requirements.