Together with Fritjof Helmchen’s group at the University of Zurich, we addressed a methodological problem that becomes increasingly important as neuroscience moves toward large-scale, multi-region recordings: how can we reliably infer statistical relationships between several neural signals at once?
Many analyses of neural data still rely on pairwise measures, such as correlations between two regions or two signals. These measures are useful, but limited: they cannot distinguish whether two sources provide unique, redundant, or synergistic information about a target. Tripartite measures — including partial correlation, variance partitioning, and partial information decomposition — are designed to address exactly this issue by decomposing multivariate relationships into more interpretable components.
However, neural recordings are never perfectly clean. They contain noise, unexplained variance, shared fluctuations, and measurement imperfections. In this work, Fomins, Sych and Helmchen systematically tested how tripartite statistical measures behave when applied to simulated neural signals with controlled levels of impurity. The results show that, although these methods can be accurate for clean signals, even modest noise can bias the estimators and lead to inflated false positive rates. Standard permutation testing, in particular, was found to be insufficiently robust in this setting.
The study therefore introduces a more conservative significance-testing procedure for tripartite measures. The goal is not to maximize detection at all costs, but to reduce the risk of reporting spurious multivariate interactions — for example, apparent redundancy or synergy — when they are actually produced by noisy or impure data. This conservative approach substantially decreases false positive rates, at the expected cost of increasing false negatives.
A useful way to summarize the contribution is: as neural datasets become richer, our statistical tools must become more careful. Detecting complex interactions between brain signals is powerful, but only if we can distinguish genuine multivariate structure from artifacts introduced by noise and imperfect measurements.
This work provides both a warning and a practical solution. It encourages researchers to treat tripartite functional relations with appropriate caution, while offering a testing framework that makes their interpretation more reliable. It is therefore a methodological contribution to the growing field of network and systems neuroscience, where understanding brain function increasingly requires going beyond pairwise connectivity toward higher-order, multivariate interactions.
To know more:
- Fomins, A., Sych, Y., & Helmchen, F. (2022). Conservative significance testing of tripartite statistical relations in multivariate neural data. Network Neuroscience, 6, 1243–1274. https://doi.org/10.1162/netn_a_00259.