NONPARAMETRIC
PERMUTATION
TESTING
CHAPTER 33
TONY YE
WHAT IS PERMUTATION
TESTING?
Framework for assessing the statistical significance of EEG
results.
Advantages:
Does not rely on distribution assumptions
Corrections for multiple comparisons are easy to incorporate
Highly appropriate for correcting multiple comparisons in EEG data
WHAT IS PARAMETRIC
STATISTICAL TESTING?
The test statistic is compared against a theoretical
distribution of test statistics expected under the H0.
t-value
2-value
Correlation coefficient
The probability (p-value) of obtaining a statistic under the H0
is at least as large as the observed statistic. [INSERT 33.1A]
NONPARAMETRIC
PERMUTATION TESTING
No assumptions are made about the theoretical underlying
distribution of test statistics under the H0.
Instead, the distribution is created from the data that you
have!
How is this done?
Shuffling condition labels over trials
Shuffling condition labels over subjects
Within-subject analyses
Group-level analyses
Recomputing the test statistic
NULL-HYPOTHESIS
DISTRIBUTION
Evaluating your hypothesis using a t-test of alpha power
between two conditions.
Two types of tests:
Discrete tests
Compare conditions
Continuous tests
Correlating two continuous variables
DISCRETE TESTS
Compare EEG activity between Condition A & B
H0 = No difference between conditions
Random relabeling of conditions
Test Statistic (TS) = as large as the TS BEFORE the
random relabeling.
Steps
1. Randomly swap condition labels from many trials
2. Compute t-test across conditions
3. If TS 0, there is sampling error or outliers
CONTINOUS TESTS
The idea:
Testing statistical significance of a correlation coefficient
Whats the difference between this and discrete tests?
TS is created by swapping data points instead of labels
SIMILARITIES
The data are not altered
The mapping of data are shuffled around.
Analysis steps:
1. Creates a H0 TS value
2. Repeats the process many MANY times
3. The MANY iterations of H0 values creates a distribution of
TS values
SIMILARITIES
Statistical evaluation entails:
Comparing the original TS value with the new distribution
of TS
Effect is not statistically significant if
The original TS does not exceed the boundaries of the
new distribution of TS
Effect IS statistically significant if
The original TS is far away from the new TS distribution
ITERATIONS
How many do you need?
~1000 for high signal-to-noise-ratio distribution
~250 500 for permutation testing at each trial, time point,
and frequency.
Why?
Less iterations = greater chance for unusually large/small
TS by chance
More iterations = estimates of H0 distribution will be
stronger with greater reliability in significance
WARNING!
More iterations = longer computation times!
DETERMINING STATISTICAL
SIGNIFICANCE
Method 1:
1. Count the number of H0 values that are more extreme
than the original TS value
Extreme = further to the right/left tail of the distribution
2. Divide the number of H0 values by the total number of
tests
PN = p-value based on the number of suprathreshold
tests
DETERMINING STATISTICAL
SIGNIFICANCE
Method 2:
1. Convert original TS to the standard deviation of the H0
distribution
2. Convert that into the p-value
Ve = original TS
Vn = vector of H0 TS
Vn bar = mean
DETERMINING STATISTICAL
SIGNIFICANCE
Method 2 contd:
Z-value p-value
Matlab function: normcdf
p-value = PZ
Advantages:
A p- and Z-value at each
pixel can be incorporated
into the H0 distribution.
THE EVER-CHANGING P-VALUE
Warning!
P-values can change each time you recompute the H0
distribution
Rest-assured if you have sufficient iterations!
The change in p-values should be tiny
THE EVER-CHANGING P-VALUE
What if you dont have sufficient iterations?
You may get p-values that CAN affect your interpretation!
THE EVER-CHANGING P-VALUE
Variability depends on:
Clean data
Results significance
MULTIPLE COMPARISONS
The Bonferroni correction is available to use!
When to use Bonferroni:
Hypothesis-driven analyses
Testing effect for 3 different electrodes in ONE timefrequency window
If there are not too many tests
If you expect robust effects
MULTIPLE COMPARISONS
When NOT to use Bonferroni correction:
Exploratory data-driven analyses
Many tests over:
Time points
Frequency bands
Electrodes
Why not?
1. Bonferroni correction assumes that the tests are
independent which many EEG results are NOT
2. The p-value will drop and hide actual effects
3. Bonferroni correction is based only on the number of
tests, instead of the information that can be found in the
tests.
PERMUTATION TESTING TO THE
RESCUE!
This framework already incorporates multiple comparison
corrections!
Unlike Bonferroni, permutation testing:
1. Corrects for information in the tests, instead of number
of tests.
2. Provides stable p-values that can detect effects
regardless of correlated data.
NONPARAMETRIC
PERMUTATION TESTING
Two methods available
1. Corrects for multiple comparisons by using the pixel to
determine the threshold
2. Corrects by using the cluster to determine threshold.
Generally, how do these methods work?
Approaches the H0 distribution at the time-frequency map
level, instead of pixel level
Reflecting information from the ENTIRE time-frequencyelectrode space
PIXEL-BASED STATISTICS
Creating a distribution that contains the pixel from each
iteration with the most extreme statistical value.
Steps:
1.
2.
3.
4.
Generate TS values under the
H0 (as previously outlined)
Store one or two pixels with the
most extreme TS values in a
matrix
For + AND effects: Define the
statistical threshold for 2.5
percentile and 97.5th percentile
For + OR effects: Define
threshold for 95th or 5th
percentile, respectively
PIXEL-BASED STATISTICS
Things to note:
A summary of the most extreme H0 TS are saved across
all pixels, at each iteration
Map-level thresholding
Single pixels can be statistically significant
Even if neighboring pixels are non-significant
Interpretability depends on experimental design and size
of time-frequency pixels
CLUSTER-BASED CORRECTION
What is a cluster?
A group of contiguously significant points in timefrequency-electrode space
Can be seen after applying a threshold with any pixel that
has a value below it set to zero
What is cluster-based correction?
Significance = enough neighboring pixels with
suprathreshold values.
Individual pixels that are significant arent really significant
CLUSTER-BASED
CORRECTIONS
Big enough?
Number of extracted frequencies
Resolution of the results
Temporally downsampled
Example:
1 time point @ 1 ms
Significance is false
1 time point @ several Hz and several hundreds of ms
Significance is valid
CLUSTER-BASED
CORRECTIONS
Non-data-driven method:
1. Predefine a number of target time and frequency points
E.g., Clusters of 200 ms, 3 Hz
2. Remove clusters that are less than that number
Data-driven method:
1. Perform permutation testing (as previously outlined)
2. At each iteration, apply a threshold to the time-frequency
map using an uncorrected p-value (precluster threshold)
3. Threshold the H0 iteration map
THRESHOLDING
STRATEGIES
Method 1:
Use parametric statistics
P-value from your t-test or correlation coefficient
Ideal for normally distributed data
Method 2:
Loop through the iterations TWICE
Once to build the H0 distribution at each pixel
Second to threshold them using nonparametric pixelbased significance thresholding.
THRESHOLDING
STRATEGIES
Now you have a distribution of the largest suprathreshold
clusters under the H0
Threshold the map of observed statistical values using
uncorrected p-value
Next steps:
1. Identify clusters in threshold map
2. Remove clusters that
are less than 95th%
of the largest cluster
distribution
CLUSTER-BASED
METHOD SUMMARY
Performs map-level thresholding
Corrections are based on the information within the data,
instead of the number of tests performed
Precluster threshold affects the cluster correction threshold!
Small p-value = remove many clusters
Larger p-value = leave many clusters
Extremely sensitive to large clusters
Localized true effects may go unnoticed!
FALSE DISCOVERY
RATE (FDR) METHOD
How does it work?
Controls for the probability of Type I errors within a distribution of
p-values
function [pID,pN] = fdr(p,q)
pID = p-value based on independence or positive dependence
pN = nonparametric p-value
p = p-value vector
q = FDR level
Limitations
Critical p-value is based on the number of tests performed AND
distribution of p-values
SUMMARY
Shuffling depends on your focus of analysis and hypothesis
Comparing two conditions
Correlations of time-frequency and reaction times over
trials
Shuffle condition labels
Shuffle mapping of reaction time to trials
Connectivity between two electrodes
Shuffle ordering of time segments
If still unsure
What effects does your hypothesis concern?
Sometimes, shuffling can be performed with more than
one option
SUMMARY
What about complex statistical designs?
There is a lack of support for complex analyses in
cognitive electrophysiology
Take a hypothesis-testing approach with SPSS, SAS, or R
first
How can you report your analyses in a methods section?
What variables were shuffled?
How many iterations?
How were p-values created?
Which multiple comparisons correction did you use?