Analyzing EEG time-frequency data in a factorial design

Question

I am trying to find the best strategy to analyze a set of EEG time-frequency data from 24 subjects, using 64 electrodes. This is an exploratory study as I have no a-priori hypothesis on where or when changes in two frequency bands might be observed. In this case I have found that using non-parametrical statistical testing might be the best soulution: in particular, cluster-based permutation testing is widely used in cognitive electrophysiology.

However this approach is not feasible as I have a 2X2X2 within subject design, and it is not possible to look at interaction with permutation testing. On the other hand, running an ANOVA on the whole data-set might well result in an high number of false positives.

A solution to my proble might be to define ROI(s), so to perform ANOVA(s) on a limited number of electrodes within a time window. However I have no clue on how to define such ROI(s).

What is the best strategy according to you?

Answer 1

It is possible to look at within-subject interactions using cluster permutation tests. I.e. Eric Maris here: https://mailman.science.ru.nl/pipermail/fieldtrip/2011-January/003447.html

2x2x2 with factors A,B,C has 4 interactions AxB, AxC, BxC and AxBxC and 8 cells

+--------+-------------+-------------+
| C = 0  |    B = 0    |    B = 1    |
+--------+-------------+-------------+
| A = 0  | A0 B0 C0    | A0 B1 C0    |
| A = 1  | A1 B0 C0    | A1 B1 C0    |
+--------+-------------+-------------+

+--------+-------------+-------------+
| C = 1  |    B = 0    |    B = 1    |
+--------+-------------+-------------+
| A = 0  | A0 B0 C1    |  A0 B1 C1   |
| A = 1  | A1 B0 C1    |  A1 B1 C1   |
+--------+-------------+-------------+

2-way: You calculate the difference of differences at the average of missing factor:

$$ interaction_{AB} = \frac{interaction_{AB_{c=0}} + interaction_{AB_{c=1}}}{2} = \frac{(A_0B_0C_0-A_1B_0C_0)- (A_0B_1C_0-A_1B_1C_0) + (A_0B_0C_1-A_1B_0C_1)- (A_0B_1C_1-A_1B_1C_1)}{2} $$

and for the three way interaction is the difference of interactions $$ interaction_{AB_{c=1}} - interaction_{AB_{c=0}} $$ (or any of the other pairs, same result)

I wrote a small script that confirms this:

library(tidyverse)
library("ANOVApower")

# simulate simple 2x2x2 design
design_result <- ANOVA_design(design = "2w*2w*2w",
                              n = 10, 
                              mu = c(1:8), 
                              sd = 1.0, 
                              labelnames = c("A",0,1,"B",0,1,"C",0,1))

d = design_result$dataframe
contrasts(d$A) = c(-.5,.5)
contrasts(d$B) = c(-.5,.5)
contrasts(d$C) = c(-.5,.5)

# Cool, the ANOVApower script already gives us a unique column for each cell
# we can forget about subject because everything is balanced :)
d_cell = d%>%group_by(cond) %>%
  summarise(y = mean(y))


d_interactions = d_cell %>% summarise(
                     AB_C0 = (y[cond=='A_0_B_0_C_0']-y[cond=='A_0_B_1_C_0']) - (y[cond=='A_1_B_0_C_0']-y[cond=='A_1_B_1_C_0']),
                     AB_C1 = (y[cond=='A_0_B_0_C_1']-y[cond=='A_0_B_1_C_1']) - (y[cond=='A_1_B_0_C_1']-y[cond=='A_1_B_1_C_1']),
                     AC_B0 = (y[cond=='A_0_B_0_C_0']-y[cond=='A_0_B_0_C_1']) - (y[cond=='A_1_B_0_C_0']-y[cond=='A_1_B_0_C_1']),
                     AC_B1 = (y[cond=='A_0_B_1_C_0']-y[cond=='A_0_B_1_C_1']) - (y[cond=='A_1_B_1_C_0']-y[cond=='A_1_B_1_C_1']),
                     CB_A0 = (y[cond=='A_0_B_0_C_0']-y[cond=='A_0_B_1_C_0']) - (y[cond=='A_0_B_0_C_1']-y[cond=='A_0_B_1_C_1']),
                     CB_A1 = (y[cond=='A_1_B_0_C_0']-y[cond=='A_1_B_1_C_0']) - (y[cond=='A_1_B_0_C_1']-y[cond=='A_1_B_1_C_1'])
)

d_2way = d_interactions %>% summarise(AB = (AB_C0 + AB_C1)/2,
                             AC = (AC_B0 + AC_B1)/2,
                             BC = (CB_A0 + CB_A1)/2)


d_interactions$AB_C0 - d_interactions$AB_C1
d_interactions$AB_C1 - d_interactions$AB_C0

# confirm with linear model results
lm(y~A*B*C,d)

Which returns:

> d_2way
# A tibble: 1 x 3
      AB     AC    BC
   <dbl>  <dbl> <dbl>
1 -0.683 -0.827 0.129

> d_interactions$AB_C1 - d_interactions$AB_C0
[1] 1.307831
> d_interactions$AB_C1 - d_interactions$AB_C0
[1] 1.307831

> lm(y~A*B*C,d)

Call:
lm(formula = y ~ A * B * C, data = d)

Coefficients:
(Intercept)           A1           B1           C1        A1:B1        A1:C1        B1:C1     A1:B1:C1  
     4.3853       3.8473       1.8056       1.1670      -0.6832      -0.8272       0.1288       1.3078  

Identical results.

This is what I would do

Answer 2

This scenario is practically unsolvable. If one doesn't have any assumptions about when, where, and how the changes may occur, then for 24 subjects (typically in this business, with high inter-subject variance due to age, sex, etc.) x 64 channels (usually, with variable SnR) x 45 frequencies (each signal exists in the frequency space of 1 to 45 Hz, at least) x some_cognitive_states_or_tasks you have an insane number of alternatives. It will take equally insane statistical power (i.e., amount of evidence) to prove any effect. The expected effect size on the field of cognitive neuroscience is below Cohen d < 1 (see, 10.1371/journal.pbio.2000797), so even if you have it somewhere, it will not be statistically significant.