How to control for confounds in decoding analyses of neuroimaging data

Area of Science:

• Neuroimaging
• Machine Learning
• Statistical Analysis

Background:

• Multivariate decoding analyses are increasingly used in neuroimaging, offering an alternative to mass-univariate methods.
• A key limitation is identifying the specific information source driving decoding performance, especially when confounds are present.
• Confounding variables can obscure the true interpretation of decoding results in neuroimaging studies.

Purpose of the Study:

• To evaluate the effectiveness of post hoc counterbalancing and confound regression in controlling for confounding variables in decoding analyses.
• To assess the biases introduced by these methods using both simulations and empirical neuroimaging data.
• To identify a reliable method for controlling confounds in decoding analyses to improve the interpretability of neuroimaging findings.

Main Methods:

• The study employed comprehensive simulations and analyses of empirical structural MRI data.
• Two methods for controlling confounds were evaluated: post hoc counterbalancing and confound regression.
• The impact of performing confound regression within each cross-validation fold was specifically investigated.

Main Results:

• Post hoc counterbalancing introduced a positive bias, inflating decoding performance.
• Standard confound regression resulted in a negative bias, sometimes yielding below-chance performance.
• Performing confound regression within each cross-validation fold eliminated bias and produced plausible decoding performance in both simulations and empirical data.

Conclusions:

• Post hoc counterbalancing and standard confound regression are unreliable for controlling confounds in decoding analyses.
• Cross-validated confound regression is a robust method for appropriately controlling for confounding variables in neuroimaging decoding.
• This validated approach enables more accurate insights into the specific information driving decoding performance in neuroimaging research.