A study by Anders Eklund, Thomas E. Nichols, and Hans Knutsson reveals that common statistical methods used in functional magnetic resonance imaging (fMRI) analyses, such as SPM, FSL, and AFNI, produce inflated false-positive rates for clusterwise inference, despite being conservative for voxelwise inference. The researchers used resting-state data and performed 3 million random task group analyses to evaluate the familywise error rates of these software packages. They found that parametric methods, which assume Gaussian spatial autocorrelation functions, fail to control false positives for cluster inference, while nonparametric permutation tests produce nominal results for both voxelwise and clusterwise inference. The study highlights the need for validating statistical methods in neuroimaging, as the validity of many published fMRI studies may be compromised. The findings suggest that spatial autocorrelation functions that do not follow the assumed Gaussian shape are a major cause of invalid cluster inferences. The nonparametric permutation test is found to be more reliable for both voxelwise and clusterwise inference. The study also shows that the most common cluster extent threshold (80 mm³) leads to high false-positive rates. The results indicate that the problems extend to task-based fMRI data as well. The study emphasizes the importance of data sharing and the need for accurate inferential tools in neuroimaging research.A study by Anders Eklund, Thomas E. Nichols, and Hans Knutsson reveals that common statistical methods used in functional magnetic resonance imaging (fMRI) analyses, such as SPM, FSL, and AFNI, produce inflated false-positive rates for clusterwise inference, despite being conservative for voxelwise inference. The researchers used resting-state data and performed 3 million random task group analyses to evaluate the familywise error rates of these software packages. They found that parametric methods, which assume Gaussian spatial autocorrelation functions, fail to control false positives for cluster inference, while nonparametric permutation tests produce nominal results for both voxelwise and clusterwise inference. The study highlights the need for validating statistical methods in neuroimaging, as the validity of many published fMRI studies may be compromised. The findings suggest that spatial autocorrelation functions that do not follow the assumed Gaussian shape are a major cause of invalid cluster inferences. The nonparametric permutation test is found to be more reliable for both voxelwise and clusterwise inference. The study also shows that the most common cluster extent threshold (80 mm³) leads to high false-positive rates. The results indicate that the problems extend to task-based fMRI data as well. The study emphasizes the importance of data sharing and the need for accurate inferential tools in neuroimaging research.