The first step in modern lesion-deficit analysis

Sir, I am grateful to Karnath and his colleague (Karnath and Smith, 2014) for a sophisticated commentary on our recent study (Mah et al., 2014); nonetheless, four aspects of their analysis may cause some readers to misapprehend our conclusions in a way that will tend to perpetuate the errors it was our original aim to correct. First, the principal reason for changing to multivariate inference is not the complex distributed functional architecture of the brain but the complex distributed structural architecture of lesions. Just as mass-univariate inference has not been an obstacle to discovering functional networks with functional MRI, so it would not have been a (major) obstacle to discovering such networks with lesions if lesions had the spatial properties of blood oxygen level-dependent response. Multivariate inference in the context of lesionmapping is not an extension to the conventional voxelwise mass-univariate method (i.e. voxel-based lesion–symptom mapping), mainly for those who wish to examine networks as well as single critical areas, but a necessity for anyone who uses vascular lesions to do any kind of anatomical inference in the brain. For while the size of the error may well be greater where the pattern of dependence follows a multi-locus, distributed network, substantial error will nonetheless still occur with single loci, as we explicitly demonstrate in our paper. We show that the size of such error is sufficient to explain, for example, the surfeit of white matter localisations now crowding the literature. Second, the large region of interest-based multivariate approach proposed by the authors (Smith et al., 2013) does not solve the problem we have identified but arguably conceals it. We currently do not have robust functional criteria for defining large regions of interest—indeed, we need lesion-deficit mapping for this in the first place—and A