Response latencies for correctly detected targets in the left and right hemifields were analysed using a 2x2 (group by hemifield) analysis of variance. Again, since the design was intended to produce performance at or near ceiling, latencies for misses and false alarms were too few to be usefully analysed.
Functional images were corrected for head motion using Decoupled Automated Rotational and Translational motion correction [55], a method that uses a k-space representation of the images to separate rotational and translational (k-space phase) components. Linear trend and baseline offset were removed from each trial separately. A thresholding method was applied to temporally averaged, motion-corrected functional images to classify voxels as brain or non-brain, and these automatically generated classifications were then manually examined and retouched. An ideal waveform was constructed, 1 for time points within each of the task periods and 0 for time points within each of the fixation periods. The first point within each of these periods was excluded to allow for haemodynamic lag. In addition, all points within and immediately following breaks in fixation were excluded. The permutation test [10] component of the AFNI software package [31] was used to generate two-tailed probabilities for task-related activation of each brain voxel.
SPM99 [34] was used to compute spatial transformations that normalised each subject's echo-planar images onto a template brain. The probability maps computed by AFNI for the analysis of task versus fixation were then imported into SPM99 as SPM{z}'s, spatially smoothed, and subjected to the identical spatial transformations. A voxelwise t-test was applied over this set of spatially normalised individual maps, producing an SPM{t} representing group activations. Tail probabilities for clusters of active voxels were then computed with reference to the distribution of expected cluster sizes [35], for each group separately and for the comparison between groups.
As a separate analysis, within each individual data set and without blurring or spatial normalisation, probability maps from the task-versus-fixation comparison were used in combination with each subject's anatomical images to draw individualised regions of interest for a left-hemifield-versus-right-hemifield attention comparison. The three regions of interest in each hemisphere consisted of five voxels in ventral occipital cortex in the neighbourhood of the middle occipitotemporal gyrus, four voxels just superior to the fundus of the intraparietal sulcus, and four voxels in superior parietal cortex. Despite variation across subjects in the extent of activation, sizes of these regions were held constant across subjects so as to ensure comparability of error variances in the subsequent analysis of variance [12]. In cases in which activations were of insufficient extent to furnish the desired number of voxels, sub-threshold tail probabilities and local anatomical features were used to guide the placement of regions. These regions were drawn on a single slice or on adjacent slices whose anterior-posterior Talairach coordinate was approximately -70mm. Regions of interest for all subjects are illustrated in Figure 1.
|
For the attention comparison, behavioural data from each subject were used to define a second ideal waveform that described the direction of attention as a function of time, within the task periods only. In order to account for haemodynamic delay the first point in each task period was excluded from this waveform, as was the first point following each shift of attention. In addition, all points during breaks in fixation were excluded. In cases of missed targets or false alarms, points were excluded backwards in time to the previous correct response or to the beginning of the task period, and forward in time to the next correct response or to the end of the task period. The points remaining after these exclusions were assigned a value of 1 for leftward attention, or 0 for rightward attention. Because this second ideal waveform was defined only within the task periods, this attention comparison was independent of the earlier comparison of task versus fixation.
The regression coefficient between the attention waveform and the fMRI time series was transformed to a z-score at each voxel, forming an SPM{z}. z-scores from this map were then averaged within each region of interest, to form a z-score reflecting the degree of attention-related activity in the group of selected voxels as a whole. Since the ideal waveform was arbitrarily chosen to be positive for leftward attention and zero for rightward attention, a positive regional z-score denotes correlation with leftward attention while a negative regional z-score denotes correlation with rightward attention. Regional z-scores were subjected to a 2x2x3 (group by hemisphere by region-of-interest) analysis of variance, including age as a covariate. Post hoc t-tests were applied as indicated by F values from this analysis.
