One task is continuous overt speech production (P), the other is passive listening comprehension (C). We have the belt data but it doesn't seem to work well. The cross-correlation between aliased respiration depth signal and global mean signal varies a lot across subjects (peak latency between -16s and +16s) probably because the delayed global signal change of one task block (30s) can carry over and accumulate in the next block (18s apart). We randomized the block orders. Some subjects had Ps and Cs almost alternatively in the order and the global respiratory effect is delayed for about 12s. Some subjects had consecutive P blocks, the global signal can be anticorrelated with respiration depth so the latency for positive correlation is about -16s. In addition, the correlation coefficient between the aliased signal and global signal is only up to 0.5 so at least half of the global artifact variance may not be accounted. The correlation coefficients between the WM/GM/CSF means and the convolved task regressor (P) in the design matrix are approximately 0.5. The correlation between these measures are about 0.97-0.98. Doe that mean using the global mean signal (GM/WM/CSF whatever) as a covariate is relatively "safe" for not generating false postives/negatives and spurious anti-correlations? We really couldn't find an alternative way to model the global artifact...
For the latency variation problem (Bright et al. 2009), maybe shifting the global signal (-1, 0, 1TR) can work around? (The assumption is different regions have the same temporal pattern but different latencies and amplitudes.)