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

**Authors:** Christian Dansereau, Yassine Benhajali, Celine Risterucci, Emilio Merlo Pich, Pierre Orban, Douglas Arnold, Pierre Bellec  
**arXiv:** [1607.03392](https://arxiv.org/abs/1607.03392)  
**Date:** 2016-07-12  
**Summary page:** https://www.summarizepaper.com/en/paper/1607.03392v3/

## Key points

- Connectivity studies using resting-state functional magnetic resonance imaging (rs-fMRI) are increasingly pooling data from multiple scanning sites.
- This approach allows for faster recruitment and larger sample sizes but introduces potential biases in connectivity measures across sites.
- Researchers aimed to measure the inter-site effect in connectivity and its impact on the detection of individual and group differences.
- Real rs-fMRI datasets from 345 young, healthy subjects across eight scanning sites were used in the study.
- Typical functional networks were consistently found at the group level across all sites, indicating reliable results.
- Inter-site effects on connectivity measures were small to moderate, with an average Cohen's effect size below 0.5 across brain connections.
- Monte-Carlo simulations based on real data showed that utilizing data from heterogeneous sites slightly decreased the ability to detect changes compared to a monosite study with a general linear model (GLM).
- It had a greater impact on prediction accuracy using a support-vector machine (SVM).
- Multisite studies in rs-fMRI are feasible as long as the sample size is sufficiently large enough to mitigate potential biases introduced by site heterogeneity.
- Pooling data from multiple sites can still yield reliable findings, but there may be some decrease in detection power and prediction accuracy compared to monosite studies.
- Large sample sizes are necessary for accurate results in multisite rs-fMRI studies.

## Summary

Connectivity studies using resting-state functional magnetic resonance imaging (rs-fMRI) are increasingly pooling data from multiple scanning sites. While this approach allows for faster recruitment and larger sample sizes, it also introduces potential biases in connectivity measures across sites. In this study, the researchers aimed to measure the inter-site effect in connectivity and its impact on the detection of individual and group differences. The study utilized real rs-fMRI datasets collected from 345 young, healthy subjects across eight scanning sites with 3T scanners and varying scanning protocols. The datasets were drawn from the 1000 functional connectome project. The researchers first demonstrated that typical functional networks were consistently found at the group level across all sites, indicating reliable results. They also found that the inter-site effects on connectivity measures were small to moderate, with an average Cohen's effect size below 0.5 across brain connections. To further evaluate the impact of multisite effects, the researchers conducted Monte-Carlo simulations based on real data. These simulations assessed the detection power in statistical tests comparing two groups (with and without a specific effect) using a general linear model (GLM), as well as prediction accuracy of group labels using a support-vector machine (SVM). As a reference, they also performed similar simulations using fMRI data collected at a single site but with an identical sample size. The simulations revealed that utilizing data from heterogeneous sites only slightly decreased the ability to detect changes compared to a monosite study with GLM. However, it had a greater impact on prediction accuracy using SVM. Overall, these findings suggest that multisite studies in rs-fMRI are feasible as long as the sample size is sufficiently large enough to mitigate potential biases introduced by site heterogeneity. In conclusion, this study provides valuable insights into the inter-site effects in connectivity measures during multisite rs-fMRI studies which demonstrate that despite some decrease in detection power and prediction accuracy compared to monosite studies, pooling data from multiple sites can still yield reliable findings. These findings have important implications for the design and interpretation of multisite rs-fMRI studies emphasizing the need for large sample sizes to ensure accurate results are obtained from such studies.