- Lecture 2 - Other
analysis (multiple regression):
- Used to determine the effect of independent variables on a single dependent variable. You create a design matrix (what we think data will look like) and we see how closely it matches. (Correlation).
- Fit your prediction to a time scale to each voxel, which then gives you a data value of each condition in each voxel.
- Perform a contrast (t-test) where you compare condition 1 to condition 2. T result will indicate the magnitude (colour coded), to see if there’s a significant difference you apply p = >0.05.
- Two approaches:
Whole brain analysis – where you look at the effect voxel by voxel.
- Advantages: they require no prior hypothesis about the brain area that you are looking at.
- Disadvantages: lose spatial resolution, means you end up with the possibility of many involved areas, with not knowing which one is most involved.
Region of interest analysis
– where you use
images to determine signal changes over time and across conditions, do that
- Advantages: they are hypothesis driven, avoid the lists of possible areas. Simple: don’t need any special software. Generalisable, easy to do a meta-analysis.
- Disadvantages: easy to miss things going on elsewhere in the brain.
- Whole brain analysis – where you look at the effect voxel by voxel.
for multiple comparison:
- With every test you do, your chance of a type 1 error becomes bigger = with every single experiment we do, we will find significance by chance that is not real. The correction is to adjust the p value.
What can we tell from fMRI data?
- Correlated: you can’t say that the region activated is the only place essential for that task, or if it’s essential at all.
- Need multiple trials, but with a clever design we can find out HOW a particular task is carried out.
Limitations of a mass univariate approach
- Spatial smoothing: lose a lot of information, and make assumptions
- Subtraction method: we make assumptions about how the brain implements processes
Can we use fMRI to reveal more fine grained
Pattern classification: you can train a machine (a pattern classifier)
to distinguish two or more conditions on a training set of fMRI data.
- To test the success of this, if the result is better than expected by chance, you can conclude that the information about the conditions is encoded in the brain region under investigation.
- Pattern classification: you can train a machine (a pattern classifier) to distinguish two or more conditions on a training set of fMRI data.
What about the
pattern classification is new?
- It goes beyond functional localisation: some of questions you can ask - what information is encoded in this brain region? Was that expected?
- Data analysis (multiple regression):