Avoiding illusory effects in representational similarity analysis: What (not) to do with the diagonal

Area of Science:

• Neuroimaging
• Cognitive Neuroscience
• Computational Neuroscience

Background:

• Representational Similarity Analysis (RSA) is a key method in neuroimaging, utilizing representational dissimilarity matrices (RDMs) for cross-domain comparisons.
• RDMs offer a unified format to compare neural data, computational models, and behavioral patterns.

Discussion:

• This study identifies two critical issues in RDM construction and comparison within RSA.
• The diagonal values of RDMs, crucial for assessing within-condition reliability of neural patterns, are often unestimated, hindering overall RDM reliability assessment.
• Excluding these diagonal values is essential when comparing RDMs, as their inclusion can lead to spurious correlations, as demonstrated by simulation.

Key Insights:

• Proper RSA methodology necessitates calculating diagonal values for RDMs to ensure reliability.
• Excluding diagonal values during RDM comparison is critical to prevent artificial correlations between unrelated datasets.
• Re-evaluation of existing studies, such as Coggan et al. (2016), suggests potential inflation of findings due to improper inclusion of diagonal values.

Outlook:

• Future RSA applications should prioritize accurate estimation and appropriate handling of RDM diagonal values.
• Standardizing RDM construction and comparison protocols will enhance the robustness and reproducibility of neuroimaging findings.
• Further research should explore the impact of RDM diagonal values on various analytical approaches in cognitive neuroscience.