Thanks again for your input, Wiktor!
We like the idea of dividing our runs into 4 parts, though we wanted to clarify the recommendation to use only the nearest half of the control blocks for each block of interest: is that because it’s important for any reason to avoid using any volume of the run in more than one of the 4 parts, which would be the case if for example we looked at “ALL of 1st A + B1 + ALL of 2nd A”, etc (where the entire second A block would also be used for B2)? Or is it mainly to keep the divided series short to avoid the issues with low frequencies?
We do have 4 different orderings of the Bs for counterbalancing across subjects (i.e. so that the B1 stimulus type is not first for every subject), but have recently discovered that our selection of orderings may be influencing our ability to study effects specific to a single B block’s stimulus type (i.e. type of facial expression). We’ve used the following 4 orderings based on a “Latin square":
1234
2143
3412
4321
We’ve found that expressions 3&4 as well as 1&2 are always more correlated with each other than the other pairs, but also always appear in adjacent blocks with this sampling scheme. As such we have started running subjects with all 24 possible orderings rather than just these 4. However, we have come to understand that these inflated correlations between adjacent blocks may actually be induced by the high pass filter, and that dividing the run as you suggest may really help take care of the issue.
Cheers,
Annchen
> On Oct 14, 2017, at 8:16 AM, Wiktor Olszowy <[log in to unmask]> wrote:
>
> Considering that the faces B1/B2/B3/B4 are employed only once, maybe you could divide the scans into 4 parts/scans w.r.t. time:
>
> 2nd half of 1st A + B1 + 1st half of 2nd A
> 2nd half of 2nd A + B2 + 1st half of 3rd A
> 2nd half of 3rd A + B3 + 1st half of 4th A
> 2nd half of 4th A + B4 + 1st half of 5th A
>
> this way you would not high-pass-filter (HPF) the data, as the resulting time series would be short. In fMRI experiments the same stimulus is often applied many times, because SNR/CNR is low, and by increasing the length of the time series you increase the degrees of freedom. But you apply each stimulus only once, so I do not think you would lose any power making 4 short scans out of the long one.
>
> Apart from the HPF issue, I would be maybe worried about the fixed ordering of the faces. This might be a confounder. However, this can depend on the objectives of the study and the differences between B1/B2/B3/B4.
>
> Best,
>
> Wiktor Olszowy
>
|