Cluster Failure Revisited: Impact of First Level Design and Data Quality on Cluster False Positive Rates

TL;DR

This study revisits the impact of first-level design choices and data quality on false positive rates in fMRI cluster inference, emphasizing data cleaning's importance and the implications for past research validity.

Contribution

It provides a comprehensive analysis of how design and data quality affect false positive rates, and demonstrates that data cleaning with ICA FIX restores nominal error rates in fMRI analysis.

Findings

ICA FIX cleaning achieves nominal false positive rates

Two-sided tests improve validity of cluster inference

At least 10% of fMRI studies used problematic methods

Abstract

Methodological research rarely generates a broad interest, yet our work on the validity of cluster inference methods for functional magnetic resonance imaging (fMRI) created intense discussion on both the minutia of our approach and its implications for the discipline. In the present work, we take on various critiques of our work and further explore the limitations of our original work. We address issues about the particular event-related designs we used, considering multiple event types and randomisation of events between subjects. We consider the lack of validity found with one-sample permutation (sign flipping) tests, investigating a number of approaches to improve the false positive control of this widely used procedure. We found that the combination of a two-sided test and cleaning the data using ICA FIX resulted in nominal false positive rates for all datasets, meaning that data cleaning is not only important for resting state fMRI, but also for task fMRI. Finally, we discuss the implications of our work on the fMRI literature as a whole, estimating that at least 10% of the fMRI studies have used the most problematic cluster inference method (P = 0.01 cluster defining threshold), and how individual studies can be interpreted in light of our findings. These additional results underscore our original conclusions, on the importance of data sharing and thorough evaluation of statistical methods on realistic null data.