¿Realmente Necesitamos Métodos de Inferencia Robustos y Alternativos para RM Cerebral?

Voxel-wise statistical inference lies at the heart of quantitative multi-modal brain imaging. The general linear model with its fixed and mixed effects formulations has been the workhorse of empirical neuroscience for both structural and functional brain assessment. Yet, the validity of estimated p-values hinges upon assumptions of Gaussian distributed errors. Inference approaches based on relaxed distributional assumptions (e.g., non-parametric, robust) have been available in the statistical community for decades. Recently, there has been renewed interest in applying these methods in medical imaging. Despite theoretically attractive behavior, relaxing Gaussian assumptions comes at the practical cost of reduced power (when Gaussian assumptions are met), increased computational complexity, and limited community support. We discuss the challenges of applying robust and alternative statistical methods to medical imaging inference, characterize the conditions under which such approaches are necessary, and present a new quantitative framework to empirically justify selection of inference methods.