lecture 10 - comparisons (multiple subjects) Flashcards
fixed effect analysis
y(n,i) = b_0 + ϵ(n,i)
- assumption: there is no across-subjects/runs variation
–> - variability of subject-specific or run-specific variations is not being considered in the fixed effects model - can be done across RUNS
–> because you just want to take the mean across runs
–> not across subjects - inflates statistics
- caution against using fixed effects analysis when there is significant variation across subjects. It suggests that while it may be appropriate to use fixed effects models to analyze data from multiple runs for the same subject, applying this approach to group data without accounting for individual differences can lead to misleading results.
- inference is about only the sample itself
mixed effects analysis
y(n,i) = b_0 + u_0(i) + ϵ(n,i)
- incorporate both fixed effects and random effects
- u_0(i) is the random effect associated with each measurement i
- considers the population from which subjects were sampled, modeling between-subject variance
–> takes uncertainty at lower levels into account - this is the appropriate analysis for GROUPS/ACROSS INDIVIDUALS analysis
- results are generalizable
–> inference to population
inter-individual differences
- there is a lot of anatomical structure variability between subjects
- for this reason, we try to warp individual brains into a common space (MNI standard brain template)
bringing individual brains to a volumetric ‘standard’ space
- linear registration of subject T1 anatomy to MNI standard brain
template - non-linear registration to MNI, compensative for differences in anatomy
–> accounts for individual anatomical differences that cannot be matched by linear transformations alone
pros & cons of bringing individual brains to a VOLUMETRIC ‘standard’ space
pros:
- easy
- forgiving
- works well in subcortex
cons:
- gyrus/sulcus patterns are warped into one another
surfaces of the cortex
- white-gray matter boundary is a white line on the inside
–> axons to neurons - pial surface is a blue line on the outside of the brain
–> outermost layer of the cortex - anything that we’re interested in in terms of cerebral cortex is going to be in between these 2 layers
- we can keep a surface more or less the same, but inflate the volume so that we can overlay brains, using the curvature as a guide (surface processing)
why depth matters
- you can project voxels to both the pial and the white matter surface. this leads to differences
- this shows that the depth of sampling matters
- this is important, since the smaller your voxels are, the more they are biased to sampling only the inside or outside of brain matter
- this is very sensitive to minor distortions and misregistrations
smoothing
- Smoothing is a common preprocessing step in fMRI data analysis, which reduces noise by averaging the signal across neighboring voxels.
- this works better on the surface compared to a slice
surface based registration
- registers the subjects’ anatomy to a standard surface, not a standard volume
- unlike volumetric registration, uses curvature/folding information
–> can use more than just curvature to align across individuals - however, there is still a lot of residual inter-individual variability in anatomical locations
what does MSMAII use for surface-based registration to align across individuals
- myelin
- task activations
- resting state connectivity
- DTI tractography
- curvature
–> precise loction of functional brain regions and sensitivities within regions still vary a lot across individuals
can we ever find a ‘common’ brain space
- easier for subcortical areas
–> more phylogenetically similar across individuals
–> clearer, unambiguous boundaries between regions - harder for the cerebral cortex
–> more variability across subjects
–> unclear what modalities are leading: a boundary in one modality (histology/structure) does not equate to a boundary in another (function)
intersubject correlation
- correlations between subjects
- in resting state, there is nothing linking two brains. therefore we need a stimulus
–> often naturalistic stimulus - with people experiencing similar things, we can link two brains together, as brain responses become more similar
inter-species correlation
- monkey-human connectivity
- now we can see how a monkey region corresponds to a human region
inter-individual differences in anatomy
- could pose a problem for inter-individual correlations
- solution: procrustes analysis aligns the pattern of activation/connectivity structure across brains (= hyperalignment)
–> analysis that can be used to align the shapes of objects (in this case, brain structures) despite differences in size, rotation, and translation.
pros and cons of surface-based processing
pros:
- across subjects alignment is more accurate for the cerebral cortex
- since the cortex is organised in maps, this format is closer to the brain’s actual organization
cons:
- requires accurate (sometimes manual) coregistration with functional data
- requires accurate (sometimes manual) reconstruction of surfaces
- doesn’t work on non-surface brain structures such as basal ganglia
- harder and requires extra steps than more standard fully volumetric pipelines
