I am also a fan of percent signal change. The calculation can be tricky, and I think it is nicely
explained in Jeanette Mumford's document.
We use percent signal change in two ways. One is for potential biomarkers where we are looking
for differences between groups (usually clinical subjects and controls) in an ROI, or a set of ROIs.
In this case, the effect sizes in one or multiple ROIs enter into an image classification scheme to
check the sensitivity and specificity of the proposed biomarker.
The other way is to check for "reasonableness" of the estimated results. For example, most
cognitive experiments should show maximal contrasts of about 1% (except in visual cortex),
hence, if estimates for a single subject are much larger than that, then the estimates are likely to
be bad. Poor estimates can arise from head motion, or sporadic breathing patterns by the subject,
or sometimes from a poor design matrix that is ill-conditioned.
The ArtRepair toolbox for SPM includes an observational histogram of the distribution of
estimated contrasts over the entire brain. The answers come out in percent signal change, using
the same scaling properties as described in Jeannette's document. We use it as a quality check
that the estimated effect size results are reasonable for each subject.
On Thu, 21 Feb 2008 18:25:16 -0800, Jeanette Mumford <[log in to unmask]> wrote:
>I'm a big fan of %-signal change and calculating it correctly (see
>http://mumford.bol.ucla.edu/perchange_guide.pdf). Primarily because as a
>statistician when people ask me to do a power analysis it would be *very*
>helpful if everybody reported %-signal change in their results since I could
>use this in my power calculation (this is what all other researchers in
>non-brain fields do). Also, if you would like to compare the results of
>your study to another study in terms of the size of the BOLD signals, this
>is impossible to do without %-signal change. The beta estimate just isn't
>meaningful without extra information.
>A proper %-signal change requires not only the beta estimate, but the height
>of the regressors that were used in your model and the mean signal. If we
>analyze the same block design data and my block regressor is twice as high
>as yours, your beta will be twice as large as mine. Unless the regressor
>heights are reported (and for an event related design, what would you use?)
>there's no way to go from the reported beta estimate to yours.
>As for you second question, I haven't used SPM in a while, but I'm fairly
>sure that like FSL it grand mean scales the data, so that on average all
>subjects have approximately the same timeseries average (in FSL this is
>about 10000). This is necessary to make the higher level analysis valid for
>precisely the reason I think you pointed out. So both the %-signal change
>and the betas are comparable across subjects.
>I must admit my writeup is geared towards FSL users (sorry), but the
>beginning of it lays out how calculating %-signal change can go wrong and
>how to avoid the problems. It also describes what to use as regressor
>height for event related designs. I'm sure the instructions I give for FSL
>data could easily be adapted to SPM. If you missed the link before here it
>is again http://mumford.bol.ucla.edu/perchange_guide.pdf
>Hope that helps out!
>On Thu, Feb 21, 2008 at 7:13 AM, Ada Leung <[log in to unmask]> wrote:
>> Dear SPMers,
>> I like to arise some questions about the concepts of percent signal change
>> and welcome answers from anyone who are interested in it.
>> (1) Why do we need to report percent signal change in addition to
>> activation maps?
>> (2) Percent signal change is the difference between betas divided by mean
>> activity and multiply by 100. Then it seems to mean that the value
>> represents a normalized brain activity for an individual. If this holds,
>> then comparison of percent signal change is a better indicator than general
>> activation map. If this is correct, that why do we need to present the
>> activation map since the result from the activation map is just a test of
>> significance difference between two betas which is not "normalized"?
>> (3) Is there a situation that there is positive signal change but no
>> activation shown in the activation map?
>> (4) What is negative percentage signal change? I have experience using
>> Marsbar to compute the percent signal change of say, 3 conditions. Then
>> there are situations that the 3 conditions come up with some showing
>> negative percent signal change. Because in that stage, there is no
>> substraction of betas but just a computation of percent signal change for
>> individual condition, then what is that negative percent signal change mean?
>> (5) Why does negaitve signal change not give deactivation in the
>> activation map?
>> (6) What research group or who first proposed the use of percent signal
>> change? Is there any reference?
>> (7) Is there any reference or suggested readings about the concepts of
>> percent signal change, its methodological issues and calculation?