A whole-brain analysis runs one test per voxel. At roughly 200,000 voxels and an
uncorrected threshold of p < 0.05, about 10,000 will pass by chance alone. The
threshold that is standard for a single test is meaningless applied 200,000 times, and
correcting for that is not optional bookkeeping — it is the difference between a result
and an artefact.
There are two families of correction and they answer different questions. Family-wise error control (Bonferroni, random field theory) bounds the probability of any false positive — conservative, appropriate when a single false claim is costly. False discovery rate control (Benjamini–Hochberg) bounds the expected proportion of discoveries that are false — more permissive, appropriate when you are generating candidates for follow-up. Cluster-based methods exploit spatial structure by testing contiguous blobs rather than isolated voxels, which is more sensitive but makes the inference about the cluster, not about any voxel inside it.
Permutation testing is the most defensible general answer: shuffle the labels, rebuild the null distribution from your own data, and avoid assuming a parametric form that neuroimaging data frequently violates. It costs compute, which is the main reason people skip it.
Related¶
Power and sample size — corrections cost power
Pre-registration — fixing the threshold in advance
Reporting guidelines — what must be stated