Comparing spatial null models for brain maps.

Area of Science:

• Neuroimaging
• Computational Neuroscience
• Statistical Analysis

Background:

• Advances in neuroimaging generate high-resolution brain maps.
• Comparing these maps requires statistical inference that respects spatial properties.
• Existing methods often fail to account for spatial autocorrelation, leading to inaccurate results.

Purpose of the Study:

• To comprehensively assess ten published null frameworks for statistical inference in neuroimaging.
• To evaluate framework performance in simulations and empirical datasets for comparing brain maps.
• To investigate the impact of data resolution and spatial autocorrelation on statistical estimates.

Main Methods:

• Controlled simulations with known ground truth to assess family-wise error rates.
• Application of ten null frameworks to two empirical neuroimaging datasets.
• Analysis of framework performance in correlating brain maps and assessing spatial distribution within partitions.

Main Results:

• Naive null models without spatial autocorrelation preservation yield inflated false positive rates.
• Spatially-constrained null models provide more conservative estimates but still exhibit variable performance.
• Data resolution and parcellation had minimal impact on null model performance.

Conclusions:

• Current statistical methods for comparing brain maps are insufficient, with naive methods being overly liberal and even advanced methods showing limitations.
• There is a critical need for developing more statistically rigorous methods for neuroimaging data analysis.
• This study provides a benchmark framework for evaluating and advancing future methods for comparing brain maps.