Cluster failure or power failure? Evaluating sensitivity in cluster-level inference.
Cluster failure or power failure? Evaluating sensitivity in cluster-level inference.
复制标题
DOI：
10.1016/j.neuroimage.2019.116468
复制
发表时间：
2020-04-01
期刊：
影响因子：
5.7
通讯作者：
Constable RT
中科院分区：
文献类型：
--
作者：
Noble S;Scheinost D;Constable RT
人类神经科学的开创性工作依赖于使用基于任务的功能磁共振成像绘制大脑功能的能力，但这些推理方法的经验有效性仍在表征中。Eklund及其同事最近的一项具有里程碑意义的研究表明，基于聚类范围的流行多重比较校正具有意想不到的低特异性（即，假阳性率高）。然而，这项研究的重点是特异性，虽然重要，是不完整的。方法的有效性还取决于其灵敏度（即，真阳性率或功效），但是对聚类校正的敏感性仍然知之甚少。在这里，我们评估了金标准非参数聚类校正的灵敏度，通过重新计算来自人类连接组项目中五个任务的真实的数据，并将结果与来自完整的“地面实况”数据集（n=480-493）的结果进行比较。重要的是，我们发现校正后的灵敏度低于许多fMRI应用的实际灵敏度。特别是对中等规模效应的敏感性（|Cohen的d| =0.5）在所有任务中平均不到20%，比没有任何校正的情况下小了大约三倍。此外，集群程度校正表现出的敏感性，是独立的效应大小的空间偏差。相比之下，校正的基础上的无障碍集群增强（TFCE）的统计大约增加了一倍的敏感性，但增加了空间的偏见。这些结果表明，到目前为止，我们只测量了激活映射文献中的冰山一角，因为我们的目标是通过基于聚类扩展的推理来限制家族错误率。有必要修改我们的做法，以提高灵敏度;因此，我们总结了一系列现代战略，以提高灵敏度，同时在未来的调查中保持可观的特异性。
Pioneering work in human neuroscience has relied on the ability to map brain function using task-based fMRI, but the empirical validity of these inferential methods is still being characterized. A recent landmark study by Eklund and colleagues showed that popular multiple comparison corrections based on cluster extent suffer from unexpectedly low specificity (i.e., high false positive rate). Yet that study’s focus on specificity, while important, is incomplete. The validity of a method depends also on its sensitivity (i.e., true positive rate or power), yet the sensitivity of cluster correction remains poorly understood. Here, we assessed the sensitivity of gold-standard nonparametric cluster correction by resampling real data from five tasks in the Human Connectome Project and comparing results with those from the full “ground truth” datasets (n=480-493). Critically, we found that sensitivity after correction is lower than may be practical for many fMRI applications. In particular, sensitivity to medium-sized effects (|Cohen’s d|=0.5) was less than 20% across tasks on average, about three times smaller than without any correction. Furthermore, cluster extent correction exhibited a spatial bias in sensitivity that was independent of effect size. In comparison, correction based on the Threshold-Free Cluster Enhancement (TFCE) statistic approximately doubled sensitivity across tasks but increased spatial bias. These results suggest that we have, until now, only measured the tip of the iceberg in the activation-mapping literature due to our goal of limiting the familywise error rate through cluster extent-based inference. There is a need to revise our practices to improve sensitivity; we therefore conclude with a list of modern strategies to boost sensitivity while maintaining respectable specificity in future investigations.
登录
查看更多内容
影响因子：
3.7
作者：
Cremers HR;Wager TD;Yarkoni T
通讯作者：
Yarkoni T
影响因子：
3.5
作者：
Gorgolewski KJ;Varoquaux G;Rivera G;Schwarz Y;Ghosh SS;Maumet C;Sochat VV;Nichols TE;Poldrack RA;Poline JB;Yarkoni T;Margulies DS
通讯作者：
Margulies DS
影响因子：
5.7
作者：
Genovese, CR;Lazar, NA;Nichols, T
通讯作者：
Nichols, T
DOI：
10.1111/j.2517-6161.1995.tb02031.x
发表时间：
1995-01-01
影响因子：
5.8
作者：
BENJAMINI, Y;HOCHBERG, Y
通讯作者：
HOCHBERG, Y
DOI：
10.1006/cbmr.1996.0014
发表时间：
1996-06-01
期刊：
COMPUTERS AND BIOMEDICAL RESEARCH
影响因子：
--
作者：
Cox, RW
通讯作者：
Cox, RW