{
"cells": [
{
"cell_type": "markdown",
"metadata": {},
"source": [
"# Multiple comparison correction (MCC)\n",
"In this notebook, we'll continue with the topic of multiple comparison correction (MCC).\n",
"\n",
"**What you'll learn**: after this lab, you'll ...\n",
"\n",
"- know the relative advantages and disadvantages of different MCC techniques\n",
"\n",
"**Estimated time needed to complete**: 1 hour
\n",
"**Credits**: This notebook is based on a blog by [Matthew Brett](https://matthew-brett.github.io/teaching/random_fields.html) and a previous Matlab-based lab by H. Steven Scholte."
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"## Why do we need MCC?\n",
"Univariate analyses of fMRI data essentially test hypotheses about your data (operationalized as contrasts between your $\\hat{\\beta}$ estimates) *for each voxel* separately. So, in practice, given that the MNI (2 mm) standard template brain contains about 260,000 voxels, you're conducting 260,000 different statistical tests! The obvious problem, here, is that some tests might turn out significant, while they in fact do not contain any (task-related) activity: the result is driven just by chance.\n",
"\n",
"As a researcher, you should strive to \"filter out\" the results which are driven by noise (*false positives*) and keep the results which are actually driven by the true effect (*true positives*) as much as possible. It turns out that the more tests you do, the larger the chance is that you will find one or more *false positives*. To deal with this, researchers often employ techniques for *multiple comparison correction* (MCC): **correcting** for the inflated chance of false positives when you have **multiple** tests (**comparisons**).\n",
"\n",
"In this tutorial, we will walk you through an example (simulated) dataset on which different MCC techniques are employed. We'll focus on how these different techniques influence the chance for finding false positives."
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"### The example\n",
"We'll work with the (simulated) group-level results of a hypothetical fMRI experiment. Suppose that the subjects in our hypothetical experiment were shown pictures of cats in the scanner, because we (the experimenters) were interested in which voxels would (de)activate significantly in reponse to these cat pictures (i.e. a contrast of the cat-picture-condition against baseline).\n",
"\n",
"An example of an image shown to the subjects:\n",
"![test](cute_cat.jpeg)\n",
"\n",
"After extensive preprocessing, we fitted first-level models in which we evaluated the cat-against-baseline contrast, in which the $t$-statistic refers to how strongly each voxel responded to the pictures of cats. After doing a proper group-level analysis, we now have a group-level $t$-statistic map, reflecting whether voxels on average (de)activated in response to the pictures of cats. "
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"