Control without cause: How covariate control biases our insights into brain architecture and pathology

Inferential analysis of normal or pathological brain imaging data - as in brain mapping or the identification of neurological imaging markers - is often controlled for secondary variables. However, a rationale for covariate control is rarely given and formal criteria to identify appropriate covariates in such complex data are lacking. We investigated the impact and adequacy of covariate control in large-scale imaging data using the example of stroke lesion-deficit mapping. In 183 stroke patients, we evaluated control for age, sex, hypertension, or lesion volume when mapping real or simulated deficits. We found that the impact of covariate control varies and can be strong, but it does not necessarily improve the precision of results. Instead, it systematically shifts results towards the inversed associations between imaging features and the covariate. This effect of covariate control can bias results and, as shown in another experiment, can even create effects out of nothing. The widespread use of covariate control in the statistical analysis of clinical brain imaging data - and, likely, other biological high-dimensional data as well - may not generally improve statistical results, but it may just change them. Therefore, covariate control constitutes a problematic degree of freedom in the analysis of brain imaging data and may often not be justified at all.