 |
 |
Well, Torben has already highlighted potential problems, there's nothing really to add. Depending on design (frequency domain of the predictors) it is going to make a huge difference whether to predict filtered data based on unfiltered or filtered predictors. For very long blocks I would assume it to be a very severe confound if only one of them was filtered, for fast event-related designs it might be unproblematic. But this is really going to depend on the design, so there's no ultimate answer or solution. I'm not sure about other toolboxes but in SPM it's the default approch to filter both data and predictors during single-subject analyses. This way you also ensure that the same filter is used, which might not be the case with your data. For example, there are different variants of the discrete cosine transform (SPM relies on a DCT for high-pass filter purpose), thus even if you used the same cut-off value there might be systematic differences between "data DCT" and "predictor DCT". So you also can't solve the issue by just forwarding the filtered data into a model with the high-pass filter enabled, as the SPM-filtered version of your filtered data might differ from the filtered predictors.