I do not understand why regressing task-related activation from voxel time series should change correlation between time series within a single condition.

I am using an N-Back working memory task with 0B and 2B conditions and I am interested in the correlation between two activated regions during the 2B condition (I use a contrast in the second level analysis [0B, 2B] = [0 1])

As far as I understand, removing task related activation during the preprocessing step should just remove a constant value from all time points that occur during the 2B condition.
Removing a constant should not change the value of correlation.

However, I find that including 0B and 2B conditions as covariates in the preprocessing step reduces the correlation between a spherical source ROI and voxels within the target ROI.

Hi Benjamin

This is a good question. Your rationale is spot-on if your data has been acquired using a sparse sampling paradigm (if so you can specify so in CONN in the Setup.Basic GUI). If, on the other hand, your data acquisition was continuous then the BOLD effects associated with your conditions cannot be simply modeled as a box-car response (1/0 blocks), but rather they typically include an hemodynamic delay, as well as some "ramping-up" and "ramping-down" effects at the beginning and after the end of your blocks, respectively, which last for a few seconds (this is typically modeled by convolving your original box-car condition effects with a canonical hrf). If unmodelled/uncorrected these effects can create spurious correlations between distant regions which are only caused by common responses to the task effects, and the potential overlap between your different condition BOLD effects may in addition result in other confounding effects when computing condition-specific connectivity measures. If your blocks are relatively long one way to avoid these issues would be to simply disregard the beginning/end of your blocks and compute connectivity measures from the "center" part of your blocks (you can do something like this in CONN by selecting "hanning weights" in your first-level analysis options), and then you would not have to worry about removing the condition main effects (since those would be basically constant as you describe over those restricted segments). This approach is nevertheless not generally feasible nor optimal for  shorter blocks, so the default behavior in CONN is rather to compute weighted correlation measures for each condition (where the weights are defined by the hrf-convolved condition timeseries) after removing these above-mentioned potentially confounding main effects for each condition. This approach is most similar to PPI analyses, where the main psychological effects (i.e. the main task effects) are also entered into your model in addition to the desired physiological*task interactions of interest (i.e. condition dependent connecticity measures).

Hope this helps clarify
Alfonso


Originally posted by Benjamin Sinclair: