Hi Zhiyuan - I think the answer to this question depends on what sort of analysis you are doing. If you are using something like ALFF or coherence, it shouldn't matter (as these operate in the frequency domain and ignore phase information). However, if you are using correlation (for functional connectivity) or ReHo, it will make a difference: the correlation between an HRF and itself time-shifted by 1.95s (about the longest lag between slice acquisitions assuming a 2s TR and 37 slices) is only .79, so two identical time-series (before sampling) in voxels from these two slices would have a correlation lower than 1. And depending on the acquisition order, this might introduce strange artifacts into the results that would be hard to detect (for instance, if your scanner goes odds-ascend-evens-descend, slices at the top of the volume were collected very close together in time, while those at the bottom were collected maximally distantly, so ReHo will have a spatial bias).