Multivariate Lesion-Behavior Mapping

Lesion-behavior mapping is commonly used to infer the macroscopic functional organization of the brain from the behavioral consequences of anatomically defined focal brain injury. It is a task conventionally assumed to be tractable with mass-univariate methods by analogy with spatial inference from functional and morphometric data. Here we demonstrate that this assumption is critically violated by the fundamental nature of pathological lesions and their causal relationship with the neural substrate. This cardinal fault does not merely limit mass-univariate inference to compositionally simple distributed patterns of neural dependence; it distorts it across all plausible patterns of dependence to a degree determined by the distributed structure of the lesioning process that mass-univariate inference definitionally simplifies or ignores. We argue that high-dimensional multivariate inference is essential to robust lesion-behavior mapping and provide an overview of the approach to creating a multivariate inferential framework optimized for the task, with attention to aspects of experimental design, algorithmic development, and quantification of fidelity.