1. Zillah Boraston and Disa Sauter 31st May 2006
PPI and SEM
Zillah Boraston and Disa Sauter
31st May 2006

2. Studying Connectivity
Early functional imaging studies focussed on functional segregation
But most cognitive processes are likely to depend on the interaction between brain areas
More focus now on how brain regions influence and interact with one another
Early functional imaging studies focussed on functional segregation.
But most cognitive processes are likely to depend on the interaction between brain areas
Therefore it is useful to study how brain regions influence and interact with one another

3. Measures of Connectivity
Functional
temporal correlation between remote neurophysiological events
(Friston, Frith, Liddle & Frackowiak, 1993)
Does not tell us how these temporal correlations in activity are mediated
Effective
The influence one neural system has over another
(Friston, Frith & Frackowiak, 1993)

4. Psycho-physiological Interactions (PPIs)
Measure effective connectivity, and how it is affected by psychological variables
Looks at how brain activity can be explained by the interaction between 2 variables
an experimental variable (eg level of attention)
activity in a particular brain area (source area)
This is done voxel-by-voxel across the entire brain

5. PPIs vs typical interactions
T2 S2
T1 S2
T2 S1
T1 S1
Attend eyes
Attend mouth
Upright face
Inverted face
Stimulus
Task

6. PPIs vs typical interactions
A typical interaction
Use General Linear Model:
Y = (T1-T2) β1 + (S1-S2) β2 + (T1-T2)(S1-S2) β3 + e
A PPI
Replace one of the variables with activity in source region
Eg for source region V1:
Y = (T1-T2) β V1 β (T1-T2)V1β e

7. PPIs – an example Investigating influence of 2 factors:
V1 activity
Attention
On activity in region V5
Measure brain activity under 2 conditions of attention
V1 activity
V5 activity
no attention
attention

8. Interpreting PPIs 2 possible ways:
Contribution of source area to target area (ie the effective connectivity) depends on experimental context
Response of target area to experimental variable depends on activity of source area V1 V5
attention V1 V5
attention
Mathematically, both are equivalent, but one may be more neurologically plausible

9. Deconvolving in PPI analysis
Interaction occurs at a neural level
In fMRI we can only measure the BOLD response
Haemodynamic deconvolution uses the BOLD response to estimate underlying neural activity
The analysis is then performed on this estimated neural activity

10. Pros and cons of PPI approach
Can look at the connectivity of the source area to the entire brain, and how it interacts with the experimental variable (eg attentional state)
But
Can only look at a single source area
Not easy with event-related data
Limited in the extent to which you can infer a causal relationship

11. Structural Equation Modelling (SEM)
Another way of measuring effective connectivity
Like PPI, looks at how effective connectivity is affected by experimental variables
PET or fMRI
Looks at covariances in activity between different brain areas (the degree to which their activity is related).
Combines these data with anatomical model of how the areas are connected to one another
Connectivity can be compared over time, or across different conditions (eg different levels of attention)

12. Steps in SEM Select regions of interest
Build model specifying how they’re connected to one another. Free parameters of model are ‘path coefficients’ – represent strength of connections
See what patterns of covariance this model predicts
Compare to observed patterns of covariance
‘goodness of fit’ of model is diff between predicted and observed patterns

13. Deciding on regions
Use existing fMRI and lesion data to identify likely areas
We know how these areas are likely to be connected from
Tracer studies in animals
Diffusion Tensor Imaging (DTI) studies in humans

14. Null and Alternative models
Usually look at how effective connectivity is affected by a variable eg attention
Observe patterns of covariance under 2 conditions (attention vs no attention)
Look at data using 2 models:
Null – doesn’t let path coefficients differ across the 2 conditions
Alternative model – does let them differ
Look at goodness of fit of both models
If alternative model has significantly better fit than null model, conclude that connectivity is different across 2 conditions

15. Non-linear extensions to SEM
Simplest form of SEM uses linear models
These assume connectivity between 2 regions doesn’t depend on activity level of these regions
SEM can be extended to include non-linear terms, eg interaction terms
Eg can see which brain areas are modulating connection along a particular path

16. Pros and cons of SEM
Unlike PPI, can look at influence of many brain areas at once
But models do not allow the strength of a connection to vary over the time series