CovBat#
CovBat extends NeuroComBat by harmonizing not only means and variances but also feature covariances across sites [1].
Standard ComBat assumes that site effects are limited to location (mean) and scale (variance) shifts. However, multi-site neuroimaging studies often exhibit more complex batch effects where the correlation structure among features differs across scanners—for example, the relationship between hippocampal volume and cortical thickness may vary systematically by scanner manufacturer.
CovBat addresses this limitation by harmonizing the full covariance structure of the data.
Method#
CovBat operates in three steps:
Apply ComBat to remove mean and variance site effects
Decompose residuals via PCA within each site to obtain site-specific PC scores
Harmonize PC score covariances to a reference site, then reconstruct features