fMRI Analyses (AX-CPT)

We conducted all fMRI analyses using AFNI, version AFNI_2011.12.21¹ and FSL, version 5.0.4 (Functional Magnetic Resonance Imaging of the Brain’s Software Library²; Jenkinson et al., 2012).

Participants’ images were preprocessed in AFNI. For each scanning run (with each run composed of 40 AXCPT trials), slice-timing correction was done with AFNI’s 3dTshift, aligning all of the slices to the first slice that was acquired by using quintic interpolation. All functional volumes were motion-corrected and realigned using cubic interpolation to the volume with the least amount of signal variability, as detected by AFNI’s 3dToutcount. The T1-weighted anatomical image was then co-registered to the same lowest-variability functional volume identified with 3dToutcount by using AFNI’s align_epi_anat.py, using a local Pearson correlation cost function (Saad et al., 2009). Next, the anatomical image was warped to MNI space, and the warps were applied to the functional volumes as well. A 128s highpass filter was applied to the time-series for each run to remove any signal drift and whiten the noise. The runs were concatenated and used as input to a general linear model (GLM).

The GLM allowed determination of parameter values for both sustained activity associated with the entire trial (state effects) and for event-related responses (effects for each stimulus). This model resulted in 8 task-related regressors (4 boxcar functions and 4 TENT functions, one per condition), and 6 movement regressors of no interest, representing the unconvolved time-series of the estimated translation and rotation in the x-, y-, and z-directions. We also included 4 drift regressors using 3dDeconvolve’s “-polort 4′ option.” The boxcar functions allowed us to independently code state effects and were 7.5 s long (equivalent to the trial length), convolved with a gamma function. The TENT functions allowed us to evaluate event-related effects for each trial type. Specifically, the TENT functions allowed us to estimate the time points within the hemodynamic response epoch—estimated as 25 s (12.5 TRs)—based on unassumed hemodynamic response shapes. Unlike the Finite Impulse Response (FIR) function which has been traditionally used in fMRI analyses, TENT has extra flexibility in that the stimuli do not have to synchronize with the TR grids. Accordingly, we were able to generate a GLM similar to Braver et al. (2009), despite differences in TR (the TR in their study was 2.5 s, while the TR in the present study was 2 s). This way, we could determine any intervention-related effects based on the general linear model used by Braver et al. to determine cue and probe- intervention effects. Thus, similar to their study, we estimated a 25-s (12.5 TR) event-related epoch for each cue-probe pair (AX, AY, BX, BY).

Notwithstanding, we accounted for differences in TR to determine hemodynamically-lagged timepoints for cue-activity and probe-activity in our study. Following a cue presentation at TR = 0 s (presented for 300 ms), the hemodynamically lagged cue-activity in Braver et al.’s study was estimated to occur at TR time points 3 (TR onset = 5 s) and 4 (TR onset = 7.5 s). These TR timepoints overlapped with our TR timepoints 3 (TR onset = 4 s) and 4 (TR onset = 6 s; see Table 1).

TR discrepancies between Braver study and present study.

In Braver et al.’s study, following a probe presentation at TR = 5.2 s (presented for 300 ms), the hemodynamically-lagged probe activity was estimated to occur at time points 5 (TR onset = 10 s) and 6 (TR onset = 12.5 s). These TR timepoints overlapped with our TR timepoints 6 (TR onset = 10 s) and 7 (TR onset = 12 s; see Table 1). Taken together, the cue-related activity was estimated to occur at TRs 3 and 4, while the probe-related activity was estimated to occur at TRs 6 and 7.

We identified 17 spherical regions of interest (ROIs), by creating 5 mm spheres around the peak coordinates of the 17 regions determined by Braver et al. (2009). These regions included the middle and right inferior frontal gyrus, the inferior and superior frontal junction, and the supplementary and premotor areas. Creating 5mm spheres around the peak coordinates from other studies is an established method for ROI analysis (Poldrack and Wager, 2006). We chose this approach both as an attempt to replicate the results of Braver et al., 2009, and to ensure that our ROI analysis was unbiased (Kriegeskorte et al., 2009).

We determined cue and probe-related activity for each participant at each of these 17 ROIs. Since timing is crucial to the recruitment of proactive and reactive control, we isolated each timepoint (i.e., 3, 4, 6, and 7) for examination.

Group differences at any of these time points in each of the 17 ROIs were tested in mixed models (where normality assumption was violated, non-linear mixed models were used). Generally, for each ROI at a given time point (i.e., time point 3, 4, 6, or 7), the generated mixed model was composed of fixed effects for group (intervention or control), time (pre or post), and group × time interaction, as well as random intercept effects for each participant. A family-discovery-rate (FDR) correction was applied across all cue-based models at each timepoint (i.e., separately across all models at time point 3 and separately across all models at time point 4). Similarly, an FDR correction was applied across all probe-based models at each timepoint (i.e., separately across all models at time point 6 and separately across all models at time point 7).

Finally, we were interested in identifying whether changes in performance on the AX-CPT task were associated with changes in patterns of activation on this task. To this end, we employed a linear regression framework to examine correlations between brain activity in ROIs found to be significant for intervention-related effects and behavioral indices of interest (i.e., related performance measures on AY/BX trials).