This is the third FEAT practical. It leads you through some of the more advanced usage and concepts in both single-session and higher-level FEAT analysis.
In the last practical we have run the group analyses with the outlier de-weighting option turned off. In this section we will look at the output that is generated when outlier de-weighting is performed. Note that outlier de-weighting can be performed for any of the mixed effects analyses, but is not applicable to a fixed effects analysis.
In the interest of time, we have provided you with the stripped down results of a group Feat analysis that has been setup and run with outlier de-weighting for you.
The experiment is looking to infer the group mean effect over a group of 18 subjects. The effect of interest is the BOLD response in the feedback phase of a decision making experiment.
Move into the group Feat directory corresponding to this lower-level contrast:
stats directory are the results of the higher level run with
outlier inference turned on. Take a look at the files in that directory:
prob_outlier1.nii.gz file is a 4D niftii file
giving the probability that each subject has outlier data on a
voxelwise basis. The
global_prob_outlier1.nii.gz file is
a 3D niftii file that indicates the proportion of the subjects
classified as outliers at each voxel (see the lecture for more
details). Note there are versions of these files for each variance
group in the analysis. In this case, there was only one variance
global_prob_outlier1 file into Fslview.
File > Add Standard to load in the MNI standard
File > Add to
load in the
prob_outlier1 file, and also load in the
filtered_func_data in the
directory. Note that the
filtered_func_data file is a 4D
niftii file of the 18 subjects' first-level effect sizes (COPEs) at
each voxel, in other words the "data" that gets used to calculate the
group average at each voxel.
prob_outlier image and select
Then de-select the
prob_outlier image so that it does not
get displayed on the spatial map. Do the same for
Navigate around the brain to see where outlier behaviour occurs, and
see how this relates to the data in the
timeseries plots and to the probability that each subject has outlier data
prob_outlier1.nii.gz timeseries plots. Which subject
is causing the most outlier behaviour?
stats directory is the file
zstats1_no_outliers. These are the z-stats resulting from
the higher level analysis run with the outlier inference turned off
(note that this would not normally be available when you have run
outlier inference - it is provided for you here to allow you to make
File->Add to load the z-stats into
Fslview for the analyses with
a look at how the zstats are changed when there are outliers detected.
We can use contrasts in FEAT to investigate different levels of stimulation. In the following simulated-data example we have 3 different stimulation heights: low, medium and high, and we want to know if there is a linear or quadratic trend relating stimulation strength to response strength in different areas of the brain.
cd ~/fsl_course_data/fmri/art2 Feat &
artdatamultistims.nii.gz(note that because the data has only 3 slices a pop-up appears to notify you that the default settings in FEAT have been changed - this does not concern us, as we are going to turn off pre-stats and registration in this analysis).
Go; wait for FEAT to finish and view the results. Identify which areas of the brain have the different types of trend AND in which directions.
We add temporal derivatives into the model to account for small delays (in either direction) between the model and the data. We can view this delay; it may be physiologically interesting. By applying the following commands you can manipulate the different parameter estimate values to get a delay image. We will use the audio-visual dataset results that we have already obtained in the first FEAT practical session:
Now use fslmaths to calculate the delay. Make sure that you mask the delay by the activation - it doesn't make sense to show delay where the signal looks nothing like the model.
Recall that adding the temporal derivative to the EV shifts the EV in time. Mathematically, this can be seen from the first order Taylor expansion:
f(t+delay) ~ f(t) + df(t)/dt *delay
f(t) is the EV at time t, and
is the first derivative of the EV at time t.
In practice, we are looking to fit the shifted EV to our FMRI data, Y:
Y = f(t+delay)*C + error
If we insert the first order Taylor expansion for
f(t+delay), then we get:
Y = f(t)*C + df(t)/dt *delay*C + error
So, if we fit the following GLM to our FMRI data,
Y = X1*A + X2*B + error
where X1 is the EV and X2 is the temporal derivative of the EV,
then we can see that
delay*C=B. Hence, we can
calculate the delay as
fslmaths pe2 -mul 100 -div pe1 -mas ../cluster_mask_zstat1 delay1
View the output: this is in units of % of TRs. It seems that there is a fairly consistent shift between the visual model and the data. Reopen the webpage report in a web browser, for example
open ~/fsl_course_data/fmri/av/fmri.feat/report_poststats.html on a Mac)
You can see this lag in the first timeseries plot as a temporal
displacement between the data plot and the partial model fit; that
partial model fit is the fit due just to EV1 (as COPE1 was
[1 0]). Now
look at the full model fit curve; because the temporal derivative
(EV2) is included in the full model fit, it has done a good job of
shifting the complete fit to better match the data.
In this section we will look for an interaction effect in the audio-visual dataset:
cd ~/fsl_course_data/fmri/av Feat &
[0 0 1]and
[0 0 -1]. Give the contrasts some helpful Titles. Press View Design and Efficiency. Study the design matrix and covariance/eigenvalue plots - you can see from the Effect required values that the interactions should be found as efficiently as the primary effects.
This section shows you how basis functions can be setup and used in FEAT. The dataset we will use is a jittered single-event experiment with 200 time points. The stimulus is heat applied for 3 seconds with an average inter-stimulus interval of 70 secs. We will only analyze one slice to allow for quick processing.
cd ~/fsl_course_data/fmri/bf Feat &
To start with we will analyze the dataset assuming a fixed Gamma HRF (no basis functions) and then compare the results with a set of the optimal linear basis functions.
is fine. You need to also make an F-contrast consisting only of ev1 (just set no. of F-contrasts to 1) so that we can compare it to what we will do later.
We will now process the same data using FMRIB's Linear Optimal Basis Set (FLOBS) and compare the results.
$FSLDIR/etc/default_flobs.flobsis used - so leave that selected. You can generate your own customised FLOBS by selecting
Utils > Make_flobson the main FEAT setup Gui (if you have time later then come back and try generating your own basis functions with this GUI and feed that into FEAT).
contrast will be actually carried out by the F1 f contrast. Why is this the correct way to do the test?
Utils > Make_flobson the main FEAT setup GUI.
One problem with using Basis functions is that we have to use f contrasts to look for a significant effect. These are inherently two-tailed, hence we can not tell the difference between "activation" and "deactivation".
An unsophisticated way to remove "deactivation", for example, is to use Contrast Masking to mask the f test with the t contrast for the largest basis function EV (C1). Rerun the analysis, but this time performing this contrast masking (this is setup from the Post-stats tab on the Feat GUI; you will want to use the Mask using (Z>0) option). How do the results with and without this contrast masking differ, and how do the results compare with the single gamma HRF now?
If you click on the F-stat time course you get the FEAT Time Series Report for the f-statistic. One useful piece of information available here (in the title of the second plot) is how many voxels passed thresholding for the f-test. If you compare the two you will see that there is a (tiny) difference in the numbers.
This is the end of FEAT session 3.