Statistical power and prediction accuracy in multisite resting-state fMRI connectivity.

Area of Science:

• Neuroscience
• Neuroimaging
• Data Science

Background:

• Multisite studies in resting-state functional magnetic resonance imaging (rs-fMRI) are common for increasing sample sizes.
• However, these studies may introduce systematic biases due to variations in scanning sites and protocols.
• Understanding and quantifying these inter-site effects is crucial for accurate data interpretation.

Purpose of the Study:

• To measure the inter-site effect in rs-fMRI connectivity.
• To evaluate the impact of multisite data pooling on detecting individual and group differences.
• To assess the feasibility of multisite rs-fMRI studies.

Main Methods:

• Utilized real multisite rs-fMRI data (N=345) from 8 sites with heterogeneous scanning protocols (3T scanners).
• Empirically demonstrated the presence and magnitude of inter-site effects on functional networks.
• Conducted Monte-Carlo simulations based on real data to compare statistical power (GLM) and prediction accuracy (SVM) between monosite and multisite studies.

Main Results:

• Typical functional networks were reliably detected at the group level across all sites.
• Inter-site effects were generally small to moderate (Cohen's d < 0.5).
• Multisite data slightly decreased detection power compared to monosite studies but the impact diminished with larger sample sizes (N=120).

Conclusions:

• Multisite rs-fMRI studies are feasible, despite potential inter-site effects.
• Adequate sample sizes are essential to overcome the limitations of data heterogeneity.
• Findings support the use of pooled multisite rs-fMRI data for robust neuroimaging research.