1. Introduction

Imaging data suffer from technical between-scanner variation that hinders comparisons of images across imaging sites, scanners and over time. This includes common imaging modalities, such as MRI, fMRI and DTI, as well as measurements derived from those modalities, for instance ROI volumes, RAVENS maps, cortical thickness measurements, connectome matrices, etc. To maximize statistical power, post-processing data harmonization is a powerful technique to remove unwanted variation when combining data across scanners and sites.

In two recent papers (harmonization of DTI data and harmonization of cortical thickness measurements) we have shown that ComBat, a popular batch-effect correction tool used in genomics, succesffuly removes inter-site technical variability while preserving inter-site biological variability. We showed that ComBat performs well for multi-site imaging studies that only have a few participants per site. We also showed that ComBat was robust to unbalanced studies, in which the biological covariate of interest is not balanced across sites.

We recommend to use the ComBat harmonization method after imaging processing before downstream statistical analyses. The ComBat harmonization requires the imaging data to be represented in a matrix where rows are the imaging features (for instance voxels, ROIs or connectome edges) and columns are the participants. For example, for voxel-level analyses, this usually requires images to be registered to a common template space.

Input and parameters

Data inputs for ComBat are:

• A data matrix. The data to harmonize. Rows are features (for instance voxels or brain regions) and columns are participants.
• A batch id vector. A vector (length should be equal to the number of columns in the data matrix) that specifies the id for the batch, site, or scanner to correct for. ComBat only accepts one batch vector. You should provide the smallest unit of the study that you believe introduces unwanted variation. For instance, for a study with 2 sites and 3 scanners (1 site with 1 scanner, 1 site with 2 scanners), the id for scanner should be used.
• Biological variables. Optional design matrix specifying biological covariates that should be protected for during the removal of scanner/site effects, such as disease status, age, gender, etc.

There are several alternative modes of running ComBat:

• parametric=FALSE: will instead use a non-parametric prior method in the empirical Bayes procedure (default uses parametric priors).
• eb=FALSE: will not run the empirical Bayes procedure, and therefore location and scale parameters are not shrunk towards common factors averaged across features. This is equivalent to running a location-and-scale correction method for each feature separately. This is particularly useful for debugging and method development.
• mean.only=TRUE: will only adjust the mean of the site effects across sites (default adjusts for mean and variance). This option is recommended for studies where the variances are expected to be different across sites due to the biology.