Jody Culham
Brain and Mind Institute
Department of Psychology
University of Western Ontario
http://www.fmri4newbies.com/
fMRI Analysis
with emphasis on the General Linear Model
Last Update: January 18, 2012
Last Course: Psychology 9223, W2010, University of Western Ontario
Part 1
Statistical Intuitions
What data do we start with
…
These #s are from an obsolete scanner.
With a modern 3T, we can get 3X the slices
•
•
•
•
12 slices * 64 voxels x 64 voxels = 49,152 voxels
Each voxel has 136 time points (volumes)
Therefore, for each run, we have 6.7 million data points
We often have several runs for each experiment
Why do we need stats?
• We could, in principle, analyze data by voxel surfing: move the cursor
over different areas and see if any of the time courses look interesting
Slice 9, Voxel 0, 0
Slice 9, Voxel 1, 0
Slice 9, Voxel 22, 7
The signal is much
higher where
there is brain,
but there’s still
noise
Even where there’s no brain, there’s
noise
Slice 9, Voxel 18, 36
Slice 9, Voxel 9, 27
Here’s a voxel that responds well
whenever there’s visual stimulation
Slice 9, Voxel 13, 41
Here’s one that responds well
whenever there’s intact objects
Here’s a couple that sort of show the
right pattern but is it “real”?
Slice 9, Voxel 14, 42
Why do we need stats?
• Clearly voxel surfing isn’t a viable option. We’d have
to do it 49,152 times in this data set and it would
require a lot of subjective decisions about whether
activation was real
• This is why we need statistics
The lies and damned lies
come in when you write the
manuscript
• Statistics:
– tell us where to look for activation that is related to our
paradigm
– help us decide how likely it is that activation is “real”
Predicted Responses
•
•
•
fMRI is based on the Blood Oxygenation Level Dependent (BOLD) response
It takes about 5 sec for the blood to catch up with the brain
We can model the predicted activation in one of two ways:
1. shift the boxcar by approximately 5 seconds (2 images x 2 seconds/image = 4 sec,
close enough)
2. convolve the boxcar with the hemodynamic response to model the shape of the true
function as well as the delay
PREDICTED ACTIVATION IN VISUAL AREA
BOXCAR
SHIFTED
CONVOLVED
WITH HRF
PREDICTED ACTIVATION IN OBJECT AREA
Types of Errors
Is the region truly active?
Yes
No
Does our stat test indicate
that the region is active?
Yes
HIT
Type II
Error
No
Type I
Error
Correct
Rejection
Slide modified from Duke course
p value:
probability of a Type I error
e.g., p <.05
“There is less than a 5%
probability that a voxel our
stats have declared as
“active” is in reality NOT
active
Statistical Approaches in a Nutshell
t-tests
• compare activation levels between two conditions
• use a time-shift to account for hemodynamic lag
correlations
• model activation and see whether any areas show a similar pattern
Fourier analysis
• Do a Fourier analysis to see if there is energy at your paradigm frequency
Fourier analysis
images from Huettel,
Song & McCarthy,
2004, Functional
Magnetic Resonance
Imaging
Effect of Thresholds
r = .80
64% of variance
p < 10-33
r = .50
25% of variance
p < .000001
r = .40
16% of variance
p < .000001
r = .24
6% of variance
p < .05
r=0
0% of variance
p<1
Complications
• Not only is it hard to determine what’s real, but there are all
sorts of statistical problems
Potential problems
What’s wrong with these data?
r = .24
6% of variance
p < .05
1.
data may be
contaminated by artifacts
(e.g., head motion,
breathing artifacts)
2.
.05 * 49,152 = 2457
“significant” voxels by
chance alone
3.
many assumptions of
statistics (adjacent voxels
uncorrelated with each
other; adjacent time
points uncorrelated with
one another) are false
The General Linear Model (GLM)
GLM definition from Huettel et al.:
• a class of statistical tests that assume that the
experimental data are composed of the linear
combination of different model factors, along with
uncorrelated noise
• Model
– statistical model
• Linear
– things add up sensibly (1+1 = 2)
• note that linearity refers to the predictors in the
model and not necessarily the BOLD signal
• General
– many simpler statistical procedures such as
correlations, t-tests and ANOVAs are subsumed by
the GLM
Benefits of the GLM
• GLM is an overarching tool that can do anything that the
simpler tests do
• allows any combination of contrasts (e.g., intact scrambled, scrambled - baseline), unlike simpler
methods (correlations, t-tests, Fourier analyses)
• allows more complex designs (e.g., factorial designs)
• allows much greater flexibility for combining data within
subjects and between subjects
• allows comparisons between groups
• allows counterbalancing orders within and between
subjects
• allows modelling of known sources of noise in the data
(e.g., error trials, head motion)
Part 2
Composition of a Voxel Time Course
A Simple Experiment
Lateral Occipital Complex
• responds when subject
views objects
Intact
Objects
Blank
Screen
TIME
One volume (12 slices) every 2 seconds for 272
seconds (4 minutes, 32 seconds)
Condition changes every 16 seconds (8 volumes)
Scrambled
Objects
What’s real?
A.
C.
B.
D.
What’s real?
• I created each of those time courses based by taking
the predictor function and adding a variable amount
of random noise
signal
=
+
noise
What’s real?
Which of the data sets below is more convincing?
Formal Statistics
•
Formal statistics are just doing what your eyeball test of significance did
– Estimate how likely it is that the signal is real given how noisy the data is
•
confidence: how likely is it that the results could occur purely due to
chance?
•
“p value” = probability value
– If “p = .03”, that means there is a .03/1 or 3% chance that the results are
bogus
•
By convention, if the probability that a result could be due to chance is
less than 5% (p < .05), we say that result is statistically significant
•
Significance depends on
– signal (differences between conditions)
– noise (other variability)
– sample size (more time points are more convincing)
Let’s create a time course for one LO voxel
We’ll begin with activation
Response to Intact Objects is 4X greater than Scrambled Objects
Then we’ll assume that our modelled activation is off
because a transient component
Our modelled activation could be off for other reasons
All of the following could lead to inaccurate models
• different shape of function
• different width of function
• different latency of function
Reminder: Variability of HRF
Intersubject variability of HRF in M1
Handwerker et al., 2004, NeuroImage
Now let’s add some variability due to head motion
…though really motion is more complex
•
Head motion can be quantified with 6 parameters given in any motion correction
algorithm
–
–
–
–
–
–
x translation
y translation
z translation
xy rotation
xz rotation
yz rotation
•
For simplicity, I’ve only included parameter one in our model
•
Head motion can lead to other problems not predictable by these parameters
Now let’s throw in a pinch of linear drift
• linear drift could arise from magnet noise (e.g., parts warm up)
or physiological noise (e.g., subject’s head sinks)
and then we’ll add a dash of low frequency noise
•
•
low frequency noise can arise from magnet noise or physiological noise (e.g.,
subject’s cycles of alertness/drowsiness)
low frequency noise would occur over a range of frequencies but for simplicity,
I’ve only included one frequency (1 cycle per run) here
– Linear drift is really just very low frequency noise
and our last ingredient… some high frequency noise
•
high frequency noise can arise from magnet noise or physiological noise (e.g.,
subject’s breathing rate and heartrate)
When we add these all together, we get a realistic time
course
Part 3
General Linear Model
Now let’s be the experimenter
•
•
•
First, we take our time course and normalize it using z scores
z = (x - mean)/SD
Alternative: You can transform the data into
normalization leads to data where % BOLD signal change.
This is usually a better approach because
– mean = zero
it’s not dependent on variance
– SD = 1
If you only pay
attention to one slide in
this lecture, it should
be the next one!!!
We create a GLM with 2 predictors
× 1
=
+
+
× 2
fMRI Signal
“our data”
=
=
Design Matrix x Betas
“what we
CAN
explain”
x
“how much of
it we CAN
explain”
+
Residuals
+
“what we
CANNOT
explain”
Statistical significance is basically a ratio of
explained to unexplained variance
Implementation of GLM in SPM
 Time
Many thanks to Øystein Bech Gadmar for
creating this figure in SPM
Intact
Predictor
•
•
•
Scrambled
Predictor
SPM represents time as going down
SPM represents predictors within the design matrix as grayscale plots (where black = low,
white = high) over time
GLM includes a constant to take care of the average activation level throughout each run
–
SPM shows this explicity (BV may not)
Effect of Beta Weights
• Adjustments to the beta weights have the effect of
raising or lowering the height of the predictor while
keeping the shape constant
Dynamic Example
The beta weight is NOT a correlation
• correlations measure goodness of fit regardless of scale
• beta weights are a measure of scale
small ß
large r
small ß
small r
large ß
large r
large ß
small r
We create a GLM with 2 predictors
when 1=2
=
+
+
when 2=0.5
fMRI Signal
“our data”
=
=
Design Matrix x Betas
“what we
CAN
explain”
x
“how much of
it we CAN
explain”
+
Residuals
+
“what we
CANNOT
explain”
Statistical significance is basically a ratio of
explained to unexplained variance
Correlated Predictors
• Where possible, avoid predictors that are highly correlated with
one another
• This is why we NEVER include a baseline predictor
– baseline predictor is almost completely correlated with the sum of
existing predictors
+
=
r = -.53
r = -.53
r = -.95
Two stimulus predictors
Baseline predictor
Which model accounts for this data?
xβ=1
+
xβ=0
OR
+
xβ=1
+
xβ=0
+
xβ=0
•
Because the predictors are highly correlated, the model is
overdetermined and you can’t tell which beta combo is best
x β = -1
Orthogonalizing Regressors
Contrasts in the GLM
• We can examine whether a single predictor is significant
(compared to the baseline)
• We can also examine whether a single predictor is
significantly greater than another predictor
Contrasts
 “balanced”
Conjunction of contrasts
• e.g., (+1 -1 0) AND (+1 0 -1)
• (Bio motion - Nonbio motion) AND (Bio motion > control)
• more rigorous than balanced contrast
• hypothetical (but not actual) conjunction p value = multiple of independent p values
• e.g., .01 x .01 = .001
A Real Voxel
•
Here’s the time course from a voxel that was significant in the +Intact Scrambled comparison
Maximizing Your Power
signal
=
As we saw earlier, the GLM is
basically comparing the amount of
signal to the amount of noise
How can we improve our stats?
• increase signal
• decrease noise
• increase sample size (keep subject in longer)
+
noise
How to Reduce Noise
• If you can’t get rid of an artifact, you can include it as a
“predictor of no interest” to soak up variance
Example: Some people
include predictors from the
outcome of motion correction
algorithms
Corollary: Never leave out
predictors for conditions
that will affect your data
(e.g., error trials)
This works best when the
motion is uncorrelated with
your paradigm (predictors
of interest)
Reducing Residuals
Part 3
Deconvolution of Event-Related Designs
Using the GLM
Convolution of Single Trials
Neuronal Activity
BOLD Signal
Haemodynamic Function
Time
Time
Slide from Matt Brown
Fast fMRI Detection
A) BOLD Signal
B) Individual Haemodynamic Components
C) 2 Predictor Curves for use with GLM (summation of B)
Slide from Matt Brown
DEconvolution of Single Trials
Neuronal Activity
BOLD Signal
Haemodynamic Function
Time
Time
Slide from Matt Brown
Deconvolution Example
• time course from 4 trials of two types (pink, blue) in a “jittered” design
Summed Activation
Single Stick Predictor
• single predictor for first volume of pink trial type
Predictors for Pink Trial Type
• set of 12 predictors for subsequent volumes of pink trial
type
• need enough predictors to cover unfolding of HRF
(depends on TR)
Predictor Matrix
• Diagonal filled with 1’s
.
.
.
Predictors for the Blue Trial Type
• set of 12 predictors for subsequent volumes of blue
trial type
Predictor x Beta Weights for Pink Trial Type
• sequence of beta weights for one trial type yields an estimate of
the average activation (including HRF)
Predictor x Beta Weights for Blue Trial Type
• height of beta weights indicates amplitude of response (higher
betas = larger response)
Linear Deconvolution
Miezen et al.
2000
•
Jittering ITI also preserves linear independence among the hemodynamic
components comprising the BOLD signal.
Fast fMRI: Estimation
Pros:
• Produces time course
• Does not assume specific shape for hemodynamic function
• Robust against trial history biases (though not immune to it)
• Compound trial types possible
Cons:
• Complicated
• Unrealistic assumptions about linearity if trials are too close in time
– BOLD is non-linear with inter-event intervals < 6 sec.
– Nonlinearity becomes severe under 2 sec.
• Sensitive to noise
Part 4
Dealing with Faulty Assumptions
What’s this #*%&ing reviewer
complaining about?!
•
Particularly if you do voxelwise stats, you have to be careful
to follow the accepted standards of the field. In the past few
years the following approaches have been recommended
by the stats mavens:
1. Correction for multiple comparisons
2. Random effects analyses
3. Correction for serial correlations
Dead Salmon
poster at Human Brain Mapping conference, 2009
• 130,000 voxels
• no correction for
multiple
comparisons
Fishy Headlines
Correction for Multiple Comparisons
With conventional probability levels (e.g., p < .05) and a huge number of
comparisons (e.g., 64 x 64 x 12 = 49,152), a lot of voxels will be
significant purely by chance
e.g., .05 * 49,152 = 2458 voxels significant due to chance
How can we avoid this?
1) Bonferroni correction
• divide desired p value by number of comparisons
Example:
desired p value: p < .05
number of voxels: 50,000
required p value: p < .05 / 50,000  p < .000001
•
•
•
quite conservative
can use less stringent values
• e.g., Brain Voyager can use the number of voxels in the cortical
surface
small volume correction: use more liberal thresholds in areas of the
brain which you expected to be active
Correction for Multiple Comparisons
2) Gaussian random field theory
• Fundamental to SPM
• If data are very smooth, then the chance of noise points passing
threshold is reduced
• Can correct for the number of “resolvable elements” (“resels”)
rather than number of voxels
Slide modified from Duke course
3) Cluster correction
•
•
•
•
falsely activated voxels should be randomly dispersed
set minimum cluster size to be large enough to make it unlikely that a
cluster of that size would occur by chance
some algorithms assume that data from adjacent voxels are
uncorrelated (not true)
some algorithms (e.g., Brain Voyager) estimate and factor in spatial
smoothness of maps
• cluster threshold may differ for different contrasts
4) Test-retest reliability
•
•
•
Perform statistical tests on each half of the data
The probability of a given voxel appearing in both purely by chance is
the square of the p value used in each half
e.g., .001 x .001 = .000001
Alternatively, use the first half to select an ROI and evaluate your
hypothesis in the second half.
5) False discovery rate (Genovese et al, 2002, NeuroImage)
Is the region truly active?
Yes
No
Yes
•
HIT
Type I
Error
No
•
“controls the proportion of rejected hypotheses that are falsely rejected”
standard p value (e.g., p < .01) means that a certain proportion of all
voxels will be significant by chance (1%)
FDR uses q value (e.g., q < .01), meaning that a certain proportion of the
“activated” (colored) voxels will be significant by chance (1%)
works in theory, though in practice, my lab hasn’t been that satisfied
Does our stat test indicate
that the region is active?
•
•
Type II
Error
Correct
Rejection
6) Poor man’s Bonferroni
•
•
•
Jack up the threshold till you get rid of the schmutz (especially in air,
ventricles, white matter)
If you have a comparison where one condition is expected to produce
much more activity than the other, turn on both tails of the comparison
Jody’s rule of thumb: “If ya can’t trust the negatives, can ya trust the
positives?”
Example: MT localizer data
Moving rings > stationary rings (orange)
Stationary rings > moving rings (blue)
Correction for Temporal Correlations
Statistical methods assume that each of our time points is independent.
In the case of fMRI, this assumption is false.
Even in a “screen saver scan”, activation in a voxel at one time is correlated
with it’s activation within ~6 sec
This fact can artificially inflate your statistical significance.
Autocorrelation function
original
To calculate the magnitude of the
problem, we can compute the
autocorrelation function
shift by 1
volume
For a voxel or ROI, correlate its time
course with itself shifted in time
shift by 2
volumes
time
If there’s no autocorrelation, function
should drop from 1 to 0 abruptly – pink
line
The points circled in yellow suggest
there is some autocorrelation, especially
at a shift of 1, called AR(1)
Plot these correlations by the degree of
shift
BV can correct for the autocorrelation
to yield revised (usually lower) p values
BEFORE
AFTER
BV Preprocessing Options
Temporal Smoothing of Data
• We have the option in our software to temporally
smooth our data (i.e., remove high temporal
frequencies)
• However, I recommended that you not use this option
• Now do you understand why?
Clarification
• correction for temporal correlations is NOT necessary
with random effects analyses, only for fixed effects
and individual subjects analysis
Collapsed Fixed Effects Models
• assume that the experimental manipulation has same effect in each
subject
• treats all data as one concatenated set with one beta per predictor
(collapsed across all subjects)
• e.g.,
Intact = 2
Scrambled = .5
• strong effect in one subject can lead to significance even when others
show weak or no effects
• you can say that effect was significant in your group of subjects but
cannot generalize to other subjects that you didn’t test
Separate Subjects Models
•
one beta per predictor per subject
• e.g.,
•
•
JC: Intact = 2.1
JC: Scrambled = 0.2
DQ: Intact = 1.5
DQ: Scrambled = 1.0
KV: Intact = 1.2
KV: Scrambled = 1.3
weights each subject equally
makes data less susceptible to effects of one rogue subject
Random Effects Analysis
•
Typical fMRI stats test whether the differences between conditions are significant
in the sample of subjects we have tested
•
Often, we want to be able to generalize to the population as a whole including all
potential subjects, not just the ones we tested
•
Random effects analyses allow you to generalize to the population you tested
underpaid graduate students in need of a few bucks!
•
Brain Voyager recommends you don’t even toy with random effects unless you’ve
got 10 or more subjects (and 50+ is best)
•
Random effects analyses can really squash your data, especially if you don’t have
many subjects. Sometimes we refer to the random effects button as the “make
my activation go away” button.
•
Though standards were lower in the early days of fMRI, today it’s virtually
impossible to publish any group voxelwise data without random effects analysis
•
You don’t have to worry about it if you’re using the ROI approach because (1)
presumably the ROI has already been well-established across multiple labs; and
(2) posthoc analyses of results in an ROI approach allow you to generalize to the
population (assuming you include individual variance)
Fixed vs. Random Effects GLM
Sample Data #1
Sample Data #2
Subject
Intact
beta
Scram
beta
Diff
Subject
Intact
beta
Scram
beta
Diff
1
4
3
1
1
4
3
1
2
2
1
1
2
2
3
-1
3
4
3
1
3
4
1
3
SUM
10
7
3
SUM
10
7
3
•
Fixed Effects GLM cannot tell the difference between these data sets because
(Intact sum - Scram sum) is the same in both cases
•
In Random Effects GLM, Data set #1 would be more likely to be significant
because all 3 subjects show a trend in the same direction (intact > scrambled),
whereas in data set #2, only 2 of 3 subjects show a difference in that direction
Strategies for Exploration vs. Publication
•
Deductive approach
–
–
–
–
•
Have a specific hypothesis/contrast planned
Run all your subjects
Run the stats as planned
Publish
Inductive approach
– Run a few subjects to see if you’re on the right track
– Spend a lot of time exploring the pilot data for
interesting patterns
– “Find the story” in the data
– You may even change the experiment, run additional
subjects, or run a follow-up experiment to chase the
story
• While you need to use rigorous corrections for publication, do not be overly
conservative when exploring pilot data or you might miss interesting trends
• Random effects analyses can be quite conservative so you may want to do
exploratory analyses with fixed effects (and then run more subjects if
needed so you can publish random effects)
Part 4
To Localize or Not to Localise?
To Localize or Not to Localise?
Neuroimagers can’t even
agree how to SPELL
localiser/localizer!
Methodological Fundamentalism
The latest review I received…
Approach #1:
Voxelwise Statistics
1. You don’t necessarily need a priori hypotheses (though
sometimes you can use less conservative stats if you have them)
2. Average all of your data together in Talairach space
3. Compare two (or more) conditions using precise statistical
procedures within every voxel of the brain. Any area that passes
a carefully determined threshold is considered real.
4. Make a list of these areas and publish it.
This is the tricky part!
Voxelwise Approach: Example
•
•
•
Malach et al., 1995, PNAS
Question: Are there areas of the human brain that are more responsive to
objects than scrambled objects
You will recognize this as what we now call an LO localizer, but Malach was the
first to identify LO
LO activation is shown in red, behind MT+
activation in green
LO (red) responds more to objects, abstract sculptures
and faces than to textures, unlike visual cortex (blue)
which responds well to all stimuli
The Danger of Voxelwise Approaches
•
•
•
This is one of two tables from a paper
Some papers publish tables of activation two pages long
How can anyone make sense of so many areas?
Source: Decety et al., 1994, Nature
Approach #2:
Region of interest (ROI) analysis
•
If you are looking at a well-established area (such as visual
cortex, motor cortex, or the lateral occipital complex), it’s fairly
easy to activate and identify the area
1. Do the stats and play with the threshold till you get something
believable in the right vicinity based on anatomical location
(e.g., sulcal landmarks) or functional location (e.g., Talairach
coordinates from prior studies)
2. Once you have found the ROI, do independent experiments,
extract the time course information and determine whether
activation differences between conditions are significant
–
Because the runs that are used to generate the area are
independent from those used to test the hypothesis, liberal
statistics (p < .05) can be used
Example of ROI Approach
Culham et al., 2003, Experimental Brain Research
Does the Lateral Occipital Complex compute object shape for grasping?
Step 1: Localize LOC
Intact
Objects
Scrambled Objects
Example of ROI Approach
Culham et al., 2003, Experimental Brain Research
Does the Lateral Occipital Complex compute object shape for grasping?
Step 2: Extract LOC data from experimental runs
Grasping
Reaching
NS
p = .35
NS
p = .31
Example of ROI Approach
Very Simple Stats
% BOLD Signal
Change
Left Hem. LOC
Subject
Extract average peak
from each subject for
each condition
Grasping
1
0.02
0.03
2
0.19
0.08
3
0.04
0.01
4
5
•
•
Reaching
0.10
NS
p = .35
1.01
Then simply do a
paired t-test to see
whether the peaks are
significantly different
between conditions
0.32
NS
p = .31
-0.27
6
0.16
0.09
7
0.19
0.12
Instead of using % BOLD Signal Change, you can use beta weights
You can also do a planned contrast in Brain Voyager using a module
called the ROI GLM
Utility of Doing Both Approaches
• We also verified the result with a voxelwise approach
Verification of no LOC
activation for grasping
> reaching even at
moderate threshold (p
< .001, uncorrected)
Example: The Danger of ROI Approaches
•
•
Example 1: LOC may be a heterogeneous area with subdivisions; ROI
analyses gloss over this
Example 2: Some experiments miss important areas (e.g., Kanwisher
et al., 1997 identified one important face processing area -- the fusiform
face area, FFA -- but did not report a second area that is a very
important part of the face processing network -- the occipital face area,
OFA -- because it was less robust and consistent than the FFA.
Comparing the two approaches
Voxelwise Analyses
• Require no prior hypotheses about areas involved
• Include entire brain
• Often neglect individual differences
• Can lose spatial resolution with intersubject averaging
• Can produce meaningless “laundry lists of areas” that are difficult
to interpret
• You have to be fairly stats-savvy and include all the appropriate
statistical corrections to be certain your activation is really
significant
• Popular in Europe
Comparing the two approaches
Region of Interest (ROI) Analyses
• Extraction of ROI data can be subjected to simple stats (no need
for multiple comparisons, autocorrelation or random effects
corrections)
• Gives you more statistical power (e.g., p < .05)
• Hypothesis-driven
• Useful when hypotheses are motivated by other techniques
(e.g., electrophysiology) in specific brain regions
• ROI is not smeared due to intersubject averaging
• Important for discriminating abutting areas (e.g., V1/V2)
• Easy to analyze and interpret
• Neglects other areas which may play a fundamental role
• If multiple ROIs need to be considered, you can spend a lot of
scan time collecting localizer data (thus limiting the time available
for experimental runs)
• Works best for reliable and robust areas with unambiguous
definitions
• Popular in North America
A Proposed Resolution
• There is no reason not to do BOTH ROI analyses and
voxelwise analyses
– ROI analyses for well-defined key regions
– Voxelwise analyses to see if other regions are also involved
• Ideally, the conclusions will not differ
• If the conclusions do differ, there may be sensible reasons
– Effect in ROI but not voxelwise
• perhaps region is highly variable in stereotaxic location between subjects
• perhaps voxelwise approach is not powerful enough
– Effect in voxelwise but not ROI
• perhaps ROI is not homogenous or is context-specific
Part 5
The War of Non-Independence
Finding the Obvious
A priori probability of getting JQKA
sequence = (1/13)4 = 1/28,561
A posteriori probability of getting JQKA
sequence = 1/1 = 100%
Non-independence error
• occurs when statistical tests performed are not independent
from the means used to select the brain region
Arguments from Vul & Kanwisher, book chapter in press
Non-independence Error
Egregious example
• Identify Area X with contrast of A > B
• Do post hoc stats showing that A is statistically higher than B
• Act surprised
More subtle example of selection bias
• Identify Area X with contrast of A > B
• Do post hoc stats showing that A is statistically higher than C and C is
statistically greater than B
Arguments from Vul &
Kanwisher, book chapter in
press
Figure from Kriegeskorte et
al., 2009, Nature
Neuroscience
Double Dipping & How to Avoid It
• Kriegeskorte et al.,
2009, Nature
Neuroscience
• surveyed 134 papers in
prestiguous journals
• 42% showed at least
one example of nonindependence error
Correlations Between Individual Subjects’
Brain Activity and Behavioral Measures
Sample of Critiqued Papers:
Eisenberg, Lieberman & Williams, 2003, Science
• measured fMRI activity during social rejection
• correlated self-reported distress with brain activity
• found r = .88 in anterior cingulate cortex, an area implicated in physical pain
perception
• concluded “rejection hurts”
social exclusion
> inclusion
“Voodoo Correlations”
The original title of the paper
was not well-received by
reviewers so it was changed
even though some people still
use the term
Voodoo
2009
• reliability of personality and emotion measures: r ~ .7
• reliability of activation in a given voxel: r ~ .7
• highest expected behavior: fMRI correlation is ~.74
• so how can we have behavior: fMRI correlations of r ~.9?!
“Voodoo Correlations”
"Notably, 53% of the surveyed studies selected voxels based on a correlation with the
behavioral individual-differences measure and then used those same data to compute a
correlation within that subset of voxels."
Vul et al., 2009, Perspectives on Psychological Science
Avoiding “Voodoo”
• Use independent means to select
region and then evaluate
correlation
• Do split-half reliability test
– WARNING: This is reassuring that the
result can be replicated in your sample
but does not demonstrate that result
generalizes to the population
Is the “voodoo” problem all that bad?
•
•
High correlations can occur in legitimately analyzed data
Did voxelwise analyses use appropriate correction for multiple
comparisons?
– then result is statistically significant regardless of specific correlation
•
Is additional data being used for
1.
–
2.
–
inference purposes?
if they pretend to provide independent support, that’s bad
presentation purposes?
alternative formats can be useful in demonstrating that data is clean (e.g., time
courses look sensible; correlations are not driven by outliers)