Have you looked at what happens when you vary the HPF? What does the voxel timeseries look like in these cases?
You should look into spm_filter, spm_dctmtx.
That cerebellum cluster seems to appear starting with your component #18, which is already several times longer than your experimental session.

________________________________________
From: SPM (Statistical Parametric Mapping) [[log in to unmask]] on behalf of Sungshin Kim [[log in to unmask]]
Sent: Wednesday, October 09, 2013 12:12 PM
To: [log in to unmask]
Subject: [SPM] Important issue on neural correlates of slow motor memory component

Dear all,

I am writing this email to ask very important issue in SPM processing.

Recently, we have identified neural correlates of motor memories with multiple time scales. We suggested a computational model to predict dynamic memory states with different time scales. For example, fast memory component updates its state and slow memory updates its state slowly with longer time constants.

We used 30 different time constants from 2 s to 96.6 hours.
In the attached figure, you can see the learning curve with model prediction from 21 subjects (Fig.a) and model outputs with 30 different time constant (Fig. b).

In Fig b, Red lines indicate fast component while blue lines indicate slow component.We used each of the memory traces as a regressor and did univariate analysis. We understand they are highly redundant, but we used each regressor for a separate analysis.

We used 3 sessions, each session has 99 trials (~11 minutes).
The entire experiment time approximate 35 minutes including 1 minute break between session.

[Inline image 1]

This figure shows the identified voxels, we also summarized corrected p-values although this figure shows uncorrected p<0.001 for explorative purpose.

[Inline image 2]

The problem that we have the identified voxels with slower time constant.
In the preprocessing, we used 128 high-pass filtering to eliminate drift effects. However, the regressor with slower time constants, e.g., 92.6 hour time constant, was still correlated with voxels in the cerebellum.
The activities in the cerebellum are significant with multiple correction.

So, it seems contradictory,..because using high-pass filtering with 128 s cut-off remove the low-frequency components of the BOLD signal, but they are still correlated with very slowly changing regressor with time constant of 96.6 hours.
(Please see the first figure b)
Could you answer why we found the activities in the cerebellum?
Obviously, it should not be artifacts,..the activities were very well localized in the cerebellar area...But, we need to explain why this is the case.
Sungshin Kim, PhD
Neuroscience Program
University of Southern California