A Comparison of Strategies for Incorporating Nuisance Variables into Predictive Neuroimaging Models

In this paper we compare two different methods for dealing with so-called nuisance variables (NV) when training models to predict clinical/psychometric scales from neuroimaging data. In the first approach, the NV are used to adjust the imaging data by 'regressing out' their contribution to the image features. In the second approach, the NV are included as additional predictors in the model with a separate kernel that controls their contribution to the prediction function. We evaluate these methods using data from an fMRI and a structural MRI study, and discuss the results and interpretation of the two modelling approaches.