Cluster-level statistical inference in fMRI datasets: The unexpected behavior of random fields in high dimensions

Area of Science:

• Neuroimaging
• Statistical analysis in neuroscience
• Functional Magnetic Resonance Imaging (fMRI)

Background:

• Identifying regional effects in fMRI requires controlling for multiple testing, with concerns raised about parametric methods yielding excessive false positives.
• Nonparametric methods are suggested as an alternative due to their theoretical control of false positive rates.
• The presence of unusually large clusters in high-dimensional data may challenge standard statistical assumptions.

Purpose of the Study:

• To investigate the performance of parametric statistical methods in fMRI analysis, particularly concerning false positive rates.
• To compare the efficacy of parametric versus nonparametric methods in detecting true findings while controlling for false positives.
• To understand the impact of large clusters in Gaussian Random Fields (GRFs) on statistical inference in fMRI.

Main Methods:

• Assessed parametric method performance on simulated 1D, 2D, and 3D Gaussian Random Fields (GRFs).
• Evaluated performance using 710 real-world, resting-state fMRI datasets.
• Compared empirical familywise error rates (FWERs) and statistical power between parametric and nonparametric approaches.

Main Results:

• Simulated and real fMRI data exhibited a small percentage of very large clusters, significantly impacting parametric FWERs (up to 65%).
• When excluding these large clusters, parametric methods showed a controlled FWER of 3.24%.
• Nonparametric methods rejected these large clusters as false positives, but this led to a substantial reduction in statistical power compared to parametric methods.

Conclusions:

• Parametric methods, particularly with nonstationary smoothness modeling and appropriate cluster-defining thresholds, effectively detect true findings with minimal false positives.
• The statistical power reduction associated with nonparametric methods does not justify re-analyzing previously published fMRI studies.
• Continued use of parametric methods is recommended, emphasizing rigorous assessment of biological plausibility for all significant findings, including large clusters.