Zhang Hui, Luo Wen-Lin, Nichols Thomas E
Department of Biostatistics, University of Michigan, Ann Arbor, Michigan 48109-2029, USA.
Hum Brain Mapp. 2006 May;27(5):442-51. doi: 10.1002/hbm.20253.
Except for purely nonparametric methods, statistical methods depend on assumptions about the distribution of the data studied. While these assumptions are easily checked for a single univariate dataset with diagnostic plots, in the massively univariate model used with functional MRI (fMRI) it is impractical to check with a massive number of plots. In previous work we have demonstrated how to diagnose model assumptions and lack-of-fit for single-subject fMRI models using a working assumption of independent errors; our work depended on images and time series of summary statistics that, when simultaneously viewed dynamically, identify problem scans and voxels. In this article we extend our previous work to account for temporal autocorrelation in single-subject models and show how analogous methods can be used on group models where multiple subjects are studied. We apply these methods to the single-subject Functional Image Analysis Contest (FIAC) data and find several anomalies, but none that appear to invalidate the results for that subject. With the group FIAC data we find one subject (and possibly two more) that demonstrate a different pattern of activity. None of our conclusions would be arrived at by simply looking at images of t statistics, demonstrating the importance of model assessment through exploration of the data and diagnosis of model assumptions.
除了纯粹的非参数方法外，统计方法依赖于有关所研究数据分布的假设。虽然使用诊断图可以轻松检查单个单变量数据集的这些假设，但在功能磁共振成像（fMRI）中使用的大规模单变量模型中，通过大量的图来检查是不切实际的。在之前的工作中，我们展示了如何使用独立误差的工作假设来诊断单受试者fMRI模型的模型假设和拟合不足；我们的工作依赖于汇总统计量的图像和时间序列，当同时动态查看时，可识别有问题的扫描和体素。在本文中，我们扩展了之前的工作，以考虑单受试者模型中的时间自相关，并展示如何将类似方法应用于研究多个受试者的组模型。我们将这些方法应用于单受试者功能图像分析竞赛（FIAC）数据，发现了几个异常情况，但没有一个似乎会使该受试者的结果无效。对于组FIAC数据，我们发现一个受试者（可能还有另外两个）表现出不同的活动模式。我们的结论都不是通过简单查看t统计量的图像得出的，这表明通过数据探索和模型假设诊断进行模型评估的重要性。
