Conservative significance testing of tripartite statistical relations in multivariate neural data

Area of Science:

• Systems Neuroscience
• Computational Neuroscience
• Neuroinformatics

Background:

• Understanding neuronal interactions is crucial in systems neuroscience.
• Pairwise measures of functional relations between neuronal signals offer limited insight into interaction specificity and synergy.
• Tripartite measures can disentangle functional relations into unique, redundant, and synergistic information components.

Purpose of the Study:

• To investigate the sensitivity of tripartite measures to noise in simulated neuronal recordings.
• To evaluate the accuracy and specificity of tripartite measures under varying noise conditions.
• To address the issue of high false positive rates in significance testing of these measures.

Main Methods:

• Application of tripartite measures (partial correlation, variance partitioning, partial information decomposition) to simulated neuronal data.
• Analysis of measure accuracy and specificity with noiseless versus noisy sources.
• Evaluation of permutation testing and development of a conservative null hypothesis for significance testing.

Main Results:

• Tripartite measures demonstrate accuracy and specificity for noiseless sources but exhibit significant bias with noisy sources.
• Permutation testing of tripartite measures yields high false positive rates, even with small noise fractions and large datasets.
• The proposed conservative null hypothesis substantially reduces false positive rates at the cost of a tolerable increase in false negative rates.

Conclusions:

• Significance testing of tripartite measures requires careful consideration due to noise-induced biases and high false positive rates.
• A conservative null hypothesis offers a more reliable approach for assessing the statistical significance of tripartite measures.
• The study provides conceptual and practical guidance to mitigate pitfalls in interpreting functional neuronal relations and their significance testing.