Multivariate pattern analyses Flashcards
How do we know if there’s a difference between our conditions?
•Most common approach: Student’s t-test, applied for each voxel
• “is condition A (e.g., congruent) different from condition B (e.g., incongruent)?” at each voxel
•Each voxel is analysed as a separate experiment (‘voxelwise’ analyses)
•Threshold on p-value to control type I error
• .001 = in 1 of 1000 cases/GLMs such an error
• E.g. 20000 voxels: for 20 voxels type I error
•Multiple comparisons problem
What happens when we do multiple comparisons?
Whole brain analysis ~30,000 tests -> high expected rate of false positives (600 voxels for 5% significance threshold) -> need to correct for multiple comparisons
What is the Boneferroni correction?
Correction for multiple comparisons
Bonferroni correction: Desired probability of a type I / number of comparisons made (= number of voxels), cog science and psych
• Too conservative for fctional imaging will have such a low alpha that never find anything => false negatives
• Correction needed for number of independent tests
• Nearby voxels: joined signal and noise during data acquisition, further correlations introduced during pre-processing (spatial smoothing)
• Thus, nearby voxels do not provide independent tests
What are 2 options for corrections for multiple comparisons?
1.Voxel-based thresholding
Family Wise Error: Estimate amount of correction necessary given smoothness of data
Divide the statistical threshold (e.g. 5%) by estimated number of independent comparisons (Bonferroni correction)
p < 0.05 means < 5% chance of any false positive in the data
False Discovery Rate (FDR) Correction: Control proportion of incorrectly rejected null hypotheses
Divide statistical threshold (e.g. 5%) by number of all significant comparisons
p < 0.05 means < 5% of the active voxels are false positives
- Cluster-extent based thresholding: Detect clusters based on the number of contiguous voxels that surpass a predetermined statistical threshold
p < 0.05 means no more than 5% of active clusters are false positive
But: Eklund et al. (2016) shows actual false positive rates are much larger
(Possible solution: threshold-free cluster enhancement (TFCE))
Basic analysis decisions: Whole brain or region of interest?
Whole brain:
• First-level analysis: Contrast maps computed per participant
• Second-level: per voxel, whether …
• contrast value is different from zero across participants
• effect is different between groups of subjects
• variation across subjects correlates with a factor/covariate that differs between subjects
• Needs each voxel to be indexing the same underlying brain region in all participants:
• Normalisation (putting the brains in the same space)
• Spatial smoothing
• Spatial-hypothesis free (exploratory)
• Have a big multiple comparisons problem(30,000 voxels = 30,000 statistical tests!)
• Correction for multiple comparisons can be severe (might miss something)
What is spatial normalisation?
Transforms individual brains to make them more similar for group analysis
Linear transforms: translation, rotation, zoom, shear (affine)
Non linear transforms: shrinking or expanding certain parts
Normalisation is not perfect
Visualisation on single subject brain is anatomically biased
Beware conclusions based on details of anatomical location.
What is realignment and smoothing?
Blood is always flowing in the living brain -> we measure how much a task changes the BOLD signal
Only ~ 1% task-related signal change and BOLD signal is noisy
Signal Processing to reduce noise:
Realignment (allows ‘averaging’ across trials)
Smoothing (‘averaging’ across nearby voxels)
Spatial normalizing (allows ‘averaging’ across group)
Temporal Filtering
What is a ROI analysis? (+ 2 types)
• If you have an a priori prediction about the region of the brain involved in a task, you can increase your sensitivity (less need for correction b/c less comparisons) by using an ROI approach
• Anatomical region-of-interest (ROI) draw an ROI around the anatomical area
- In group studies often probabilistic brain atlas (MNI coordinates)
• Functional ROI Look in larger area; is this region of brain doing smtg in my exp?
- Combine anatomical criteria with functional contrast (e.g. faces>objects)
- Using criteria independent of the effects of interest, or independent data: Localiser runs
• Analyse your effect of interest ONLY in that region
• Reduces number of comparisons (average across all the voxels in that region for the comparison
• (Alternative: like whole-brain but in smaller area so have less correction for multiple comparisons = small volume correction)
Whole brain vs ROI
Whole brain
• Requires normalisation
• Spatial-hypothesis free (exploratory)
• Multiple comparisons problem (30,000 voxels = 30,000 statistical tests!)
• Correction for multiple comparisons can be severe (might miss something)
Region of Interest (ROI)
• Can be done without normalisation can be done based on individual brains so dont need normalisation
• Requires a priori hypothesis about where to look
• No information about the rest of the brain (might miss something)
• ROIs must be selected independently from main data set
What is double dipping and circular analyses?
•Decisions during analysis should be based upon criteria set prior to seeing the effect of the decisions on the results Have to make analysis decisions before looking at data
• Standard operating procedure helps
• Be transparent about decisions and how they were made
• E.g., moving more than size of one voxel à remove data
•This caveat is not specific to brain imaging
•The noisier the data and the smaller the effect size of the true effect, the more circularity affects outcome of analyses
•Double-dipping: can create significant result out of pure noise-> selection not independent of the measure itself
What statistical inferences can we make? (2 methods)
•Significant effect in set of voxels: what does this mean? -> Depends upon design
•Forward inference:
- Possible if conditions differ in only one process Only changed colour so should find brain regions interested with colour
- If process X is manipulated, region R is activated –> activation of region R is related to process X
•Reverse inference:
- When people do mental task X, they activate region Y known to be involved in mental process Z. Therefore, task X invokes Z. Interpret they’re result base on other studies that found those results
- Problematic when region is involved in > 1 cognitive process!
What does the strength of reverse inference arguments depend on (2) and 3 solutions?
a)The specificity of activation of region Y: it is only activated by Z?
b)Is it really the “same region” that is activated for X and Z?
Solutions:
• Carefully establish specificity of regions
• Establish what regions are involved in multiple tasks
• Compare activity for different tasks within same participants
What is important to remember about fMRI?
fMRI is based on an indirect measure of neural activity and we correlate this measure with task conditions
Correlation is not causation! We cannot be sure that observed activity is necessary for a specific cognitive process (e.g., could be by-product because of coupled blood supply).
If you correlate enough things together you’ll find some that correlate well
What is MVPA?
• Multi-voxel or multivariate => Compared to voxel-wise or univariate, in which nearby voxels are treated individually or expected to show similar signal (spatial smoothing)
• MVPA uses pattern across voxels
Use multipe voxels to look for patterns of activation to see how things are coded
What is correlational MVPA?
•Test whether withincondition correlation between datasets is higher than between-condition correlation
