If only the probability that Null Hypothesis is less than 1/20, the experiment
is of scientific interest.
If we treat each voxel as independent test of the hypothesis, there will be
sure chances to see some example that follows null hypothesis.
If there are 20 independent experiments, then the probability of finding a
false positive is 0.64
For 200, it’s 0.99.
Now if we are taking 20k voxels each treated as independent experiments, we
are surely to find some voxels that will pass as null hypothesis.
All these values are based on the fact that probability that sample is not a
false positive is 0.05 or 1 out of 20.
This occurs when we want to find region or voxel that activates to certain
stimuli : if we don’t take all the voxels though, this effect of false
positive will be pretty low but doing a full brain analysis, this effect might
be pretty huge.
Family wise Error [FWE]
Dead Fish Experiment : Showed social vs non-social movies. Then comparing
between voxels, they found a voxel that could be used to distinguish between
the social and non-social movies. That was to show the problem with the test.
The independent tests of a same hypothesis form a family.
On doing experiments, we have to know which experiments are interchangeable so
that we know if they belong to a family or not.
If you have different experiments running with different hypothesis, they
don’t affect each other as we discussed in the previous paragraph.
Compound relationships can be used to relate variables since mostly the
variables are not independent to each other.
Solving Problem of independent tests
Assume each voxel as an independent test
As spatial resolution increased , the test got worse.
Parametric
RFT [Random Field Theory]
Voxelwise
Clusterwise : If we have only single voxel that activates to the
signal, it is pretty likely that it’s noise. So, we look at
neighbouring voxels to see a cluster of activation in the brain. The
parameters are CDT (Cluster Defining Threshold), Number of contiguous
voxels passing that threshold. But it is WRONG. The smoothness of BLOD
signal appearing in the brain was wrong that was assumed in this
experiment.
Non-parametric test
For contrast values images, multiply half of images [entire images] with 1
or -1; then each contrast-value distribution for each voxel has zero mean
since half of the values are flipped to the other side of distribution.
This is done for each image from each subject so that the corresponding
contrast image is never changed. The relationship across subjects is
changed but not the spatial information of the contrast-image map. Now, we
get the t-value. Note the max t-value. Now, repeat for like 1000 times,
and you get a histogram of t-values. Now, we get empirical threshold for
choosing an active voxel. Choosing this threshold, it automatically makes
it so that you get a false positive 1/20 time. This is only good for
second level analysis.