Statistical and computational tools

Most fMRI studies employ a massively univariate processing strategy that relies on the subtractive framework. In the past few years, we have been developing tools to investigate how the trial-to-trial variability of response magnitude is correlated with a behavioral variable. For instance, in one study, we employed logistic regression to predict whether subjects would be correct or incorrect in a difficult working memory task (Neuron, 2002). In another investigation, we extended methods of signal detection theory utilized in monkey physiology to quantify the link between single-trial responses and the perceptual decision that a near-threshold fearful face was shown (PNAS, 2005). While these studies attempted to go beyond the purely subtractive processing strategy typically adopted in fMRI, they were still univariate strategies. More recently, we have begun investigating how techniques such as multiple logistic regression and machine learning can provide a truly multivariate processing framework in which voxel-wise patterns of activation can be linked to complex behaviors.