Interpretability of statistical approaches in speech and language neuroscience

Traditional theoretical models conceive the neural system of speech and language as a set of hierarchical modules that transform a continuous acoustic stream into discrete concepts. This modular and hierarchical view arises from traditional neuropsychology and has largely been backed up by statistical models that allow for controlled variation of a few experimental factors at a time, thus allowing clear interpretations to be made. Recently, the exploration of large datasets has led to the emergence of more complex statistical models that can capture neural patterns distributed across space and time. However, the interpretation of these models is more challenging due to increased correlations and spatio-temporal dependencies between variables, which obscure the links between neural activations and linguistic functions. To guide the experimenter and data analyst through the complexity of approaches in language neuroscience, we have designed a taxonomy that delineates the trade-off between model complexity and interpretability.