How to control for confounds in decoding analyses of neuroimaging data

摘要: Over the past decade, multivariate pattern analyses and especially decoding have become a popular alternative to traditional mass-univariate in neuroimaging research. However, fundamental limitation of is that source information driving decoder ambiguous, which becomes problematic when to-be-decoded variable confounded by variables are not primary interest. In this study, we use comprehensive set simulations empirical data evaluate two techniques were previously proposed used control for confounding analyses: counterbalancing confound regression. For our analyses, attempt decode gender from structural MRI controlling "brain size". We show both methods introduce strong biases performance: leads better performance than expected (i.e., positive bias), due subsampling process tends remove samples hard classify; regression, on other hand, worse negative even resulting significant below-chance some scenarios. simulations, accuracy can be predicted variance distribution correlations between features target. Importantly, bias disappears regression procedure performed every fold cross-validation routine, yielding plausible model performance. From these results, conclude foldwise only method appropriately controls confounds, thus gain more insight into exact source(s) one9s analysis.