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