Title: Tor D' Wager
1Functional connectivityand path models
- Tor D. Wager
- Columbia University
2Functional Connectivity
Human brain mapping has been primarily used to
provide maps that show which regions of the brain
are activated by specific tasks.
Recently, there has been an increased interest in
augmenting this type of analysis by functional
connectivity studies which describe how the
various regions interact and how these
interactions depend on experimental conditions.
3A network?
- Current fashion to call any set of regions
activated in a task a network - But what does it mean to be a network?
- Set of interconnected regions information
transfer among regions
4Connectivity
Functional Connectivity Undirected
association between two or more fMRI time series
Effective Connectivity Directed influence of
one brain region on the physiological activity
recorded in other brain regions
5Functional Connectivity
Functional connectivity analysis is usually
performed using data-driven methods which make
few assumptions about the underlying biology.
- Methods
- Seed analyses
- Psychophysiological interaction analyses
- Eigenimage analysis (PCA)
- Independent Components Analysis
- Partial Least Squares
6Effective Connectivity
Effective connectivity analysis is performed
using statistical models which make anatomically
motivated assumptions and restricts inference to
networks comprising of a number of pre-selected
regions of interest.
These methods are hypothesis driven rather than
data-driven and most applicable when it is
possible to specify a complete set of the
relevant functional areas.
- Methods
- Structural Equation Modeling
- Dynamic Causal Modeling
7Connectivity Levels of analysis
8Levels of analysisTypes of inference
9Levels of analysisTypes of inference
10Types of connectivity
- Simple functional connectivity
- Two regions, A B Seed analysis
- Pathways mediation
- Interactions moderation/modulation
- psychophysiological interactions
- Granger causality
- More complex structural equation models DCM
11Functional connectivity
A
B
- Simple functional connectivity
- Region A is correlated with Region B
- Provides information about relationships among
regions - Can specify a seed region and search for voxels
with correlated data - Can be performed on timeseries data within a
subject, or individual differences (contrast
maps, one per subject)
12Timeseries connectivity Simple analysis strategy
Brain
Heart rate
r
Z
Subj 1
r
Z
Subj 2
r
Z
Subj 3
13Correlations between brain activity and
heart-rate increases
Timeseries connectivity analysis Does VMPFC
correlate with heart rate? Yes VMPFC is the
strongest positively-correlated region.
Time (TRs, 2 s)
Threshold p lt .005
Display subjects randomly with respect to
brain-heart connectivity
Average within-subject correlation (r)
Six subjects out of 24 selected at random for
display
14Issue with timeseries connectivity
- Different hemodynamic lag in different regions
- Timeseries from different regions may not match
up, even if neural activity pattern does match up - If lags are estimated from data, temporal order
may be caused by vascular (uninteresting) or
neural (interesting) response
15Temporal relationship of VMPFC and heart
Are VMPFC changes associated with autonomic
control or feedback? VMPFC changes precede
changes in heart rate, implicating it in control.
Pos Heart earlier
Neg Brain earlier
Regions showing significant lag
Blue Brain precedes heart Red Heart precedes
brain
MPFC precedes heart autonomic control
16Mediation
A
B
M
- Mediation (Baron Kenny, 1986 Shrout Bolger,
2003) - The relationship between regions A and B is
mediated by M - Can identify functional pathways spanning gt 2
regions - Can be performed on timeseries data within a
subject, or individual differences (contrast
maps, one per subject) - Also Test of whether task-related activations in
B are mediated, or explained, by M.
Task
B
M
17Moderation/Modulation
A
M
B
- Moderation (Baron Kenny, 1986)
- The relationship between regions A and B is
moderated by M - Connectivity between A and B depends on state
(level) of M - Can be performed on timeseries data within a
subject, or individual differences (contrast
maps, one per subject) - M can be task state or other variable (e.g.,
prefrontal cortex, PFC) - In SPM, on timeseries data Psychophysiological
interaction (PPI)
18Timeseries connectivity Simple analysis strategy
Brain
Heart rate
r
Z
Subj 1
r
Z
Subj 2
r
Z
Subj 3
Task modulation of connectivity brain x task
state (green - red, rest - speech prep) For each
subject bB,H green - bB,H red
19Task-modulated connectivity
Are VMPFC - Heart Rate correlations stronger
during speech preparation than baseline?
Red Stronger correlation during speech
T-value for Speech - Baseline
20Granger Causality
A
B
- Directional connectivity model
- Activity in A predicts subsequent activity in B
- Used on timeseries data
- Good for handling correlations when A and B have
different hysteresis (impulse response functions)
- Does not itself provide evidence for causality.
- Rooster crowing may Granger cause the sun to
rise - Implemented in BrainVoyager software
21Structural Equation Models
A
D
B
E
C
- Confirmatory model specification and comparison
- Allows a priori specification of regions and
relationships based on anatomical data - Models total covariance matrix among variables
using smaller set of modeled relationships - Can compare nested models or test goodness of fit
of a model - Good for single subjects, difficult to make group
inferences
22Functional vs. EffectiveScope of inference
- A goal of functional connectivity analysis is to
make inferences on the structure of relationships
among brain regions - These regions form a network
- Regions are more connected during task A than
B - This task is associated with activation of pain
pathways - A goal of effective connectivity analysis is to
make statements about causal effects among tasks
and regions. - Frontal cortex enhances connectivity between
visual areas and hippocampus. - VMPFC inhibits the amygdala
23Functional vs. EffectiveScope of inference
- A goal of functional connectivity analysis is to
make inferences on the structure of relationships
among brain regions - A goal of effective connectivity analysis is to
make statements about causal effects among tasks
and regions. - Effective connectivity provides more
theoretically powerful inference, but much
stronger assumptions! - Validity of causal inference depends strongly on
assumptions being correct. - Data is the same in either case massively
observational. - Assumptions are often poorly specified and hard
to check - !
24To claim causal effect
- You must be willing to claim that there are no
unmodeled or improperly modeled variables - Relationships are linear
- No other brain regions that are correlated with
those in model - If observational data You must be willing to
assume that arrows in model capture correct
direction of causality, based on other sources of
evidence
25Example
A
B
- If A and B are brain regions in an fMRI study,
the data is that A and B are correlated. - If I make a claim about A causing B, I must
assume - There is no other variable C that influences both
A and B - B does not cause A
26Example II
b2
A
B
b1
C
- In some cases, I may be able to estimate
coefficients for effects of A on B and B on A. - I must assume that there is some variable C that
affects B, but not A. Note the strong assumption!
- In this case, any observed association between C
and A is attributable to b2, whereas the
observed association between A and B is
determined by both b1 and b2.
27Causal inference
- Inferences about causality are very dicey when
made from observational data (e.g., Rubin) --
e.g., inferences about effects of one region on
another - Even anatomical pathways are massively
bidirectional, so little help from anatomy. - Causal relationships can be inferred for some
effects, with careful experimental design.
Properly randomized experimental treatments can
be said to have causal effects on brain
activity.
28The end.
29Example
A
B
- If A and B are brain regions in an fMRI study,
the data is that A and B are correlated. - If I make a claim about A causing B, I must
assume - There is no other variable C that influences both
A and B - B does not cause A
30- The brain pathway is a natural unit of analysis
- Multiple levels, with different interpretations
- Each is useful because they have complementary
problems - Need for complementary approaches
- Testing relationships among defined areas
- Pathway-building when all exact regions are
unknown
31Connectivity
32Brain-heart connectivity
Activation t-value
Now how is the network differ for resilient
subjects?
33Resilience effects in VMPFC
Significant (p lt .05)
Resilience
Onset estimates
Observed t
Other sig. effects Nonresilient gt Resilient in
DLPFC Resilient gt Non in ventral striatum
34Example
A
B
- If A and B are brain regions in an fMRI study,
the data is that A and B are correlated. - If I make a claim about A causing B, I must
assume - There is no other variable C that influences both
A and B - B does not cause A