Empirical analyses of BOLD fMRI statistics. II. Spatially smoothed data collected under null-hypothesis and experimental conditions

Abstract

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.