Aguirre G K, Zarahn E, D'Esposito M
Department of Neurology, Hospital of the University of Pennsylvania, Philadelphia 19104-4283, USA.
Neuroimage. 1997 Apr;5(3):199-212. doi: 10.1006/nimg.1997.0264.
In the companion to this paper (E. Zarahn, G. K. Aguirre, and M. D'Esposito, 1997, NeuroImage, 179-197), we describe an implementation of a general linear model for autocorrelated observations in which the voxel-wise false-positive rates in fMRI "noise" datasets were stabilized and brought close to theoretical values. Here, implementations of the model are tested for use with statistical parametric mapping analysis of spatially smoothed fMRI data. Analyses using varying models of intrinsic temporal autocorrelation and either including or excluding a global signal covariate were conducted upon human subject data collected under null hypothesis as well as under experimental conditions. We found that smoothing with an empirically derived impulse response function (IRF), combined with a model of the intrinsic temporal autocorrelation in spatially smoothed fMRI data, resulted in a map-wise false-positive rate which did not exceed a 5% level when a nominal alpha = 0.05 tabular threshold was applied. Use of other models of intrinsic temporal autocorrelation resulted in map-wise false-positive rates that significantly exceeded this level. fMRI data collected while subjects performed a behavioral task were used to examine (a) task-dependent global signal changes and (b) the dependence of sensitivity on the temporal smoothing kernel and inclusion/exclusion of a global signal covariate. The global signal changes within an fMRI dataset were shown to be influenced by the performance of a behavioral task. However, the inclusion of this measure as a covariate did not have an adverse affect upon our measure of sensitivity. Finally, use of an empirically derived estimate of the IRF of the system was shown to result in greater map-wise sensitivity for signal changes than the use of a broader (in time) Poisson (parameter = 8 s) kernel.
在本文的配套论文中（E. 扎拉恩、G. K. 阿吉雷和M. 德埃斯波西托，1997年，《神经影像学》，第179 - 197页），我们描述了一种针对自相关观测值的通用线性模型的实现方式，其中功能磁共振成像（fMRI）“噪声”数据集中的体素水平假阳性率得以稳定，并接近理论值。在此，对该模型的实现方式进行测试，以用于对空间平滑后的fMRI数据进行统计参数映射分析。使用不同的内在时间自相关模型，并在零假设以及实验条件下收集的人类受试者数据上进行分析，分析中既包括也排除了全局信号协变量。我们发现，使用根据经验得出的脉冲响应函数（IRF）进行平滑处理，并结合空间平滑后的fMRI数据中的内在时间自相关模型，当应用名义α = 0.05的表格阈值时，得到的全图假阳性率不超过5%水平。使用其他内在时间自相关模型会导致全图假阳性率显著超过该水平。在受试者执行行为任务时收集的fMRI数据用于检验：（a）任务相关的全局信号变化，以及（b）敏感性对时间平滑核以及全局信号协变量的包含/排除的依赖性。结果表明，fMRI数据集中的全局信号变化受行为任务表现的影响。然而，将该测量值作为协变量纳入对我们的敏感性测量并无不利影响。最后，结果表明，使用根据经验得出的系统IRF估计值比使用更宽（在时间上）的泊松（参数 = 8秒）核在全图信号变化敏感性方面更高。
