Stronger inference with direct manipulation of brain function

Abstract

At an early stage in our educations we have all learned that statistical significance need not imply functional significance, and, relatedly, that correlation does not imply causation. Despite these caveats, it is indisputable that cognitive neuroscience has learned a great deal about the neural bases of behaviour via the fundamentally correlative methods of functional neuroimaging. This is particularly true in cases when a carefully controlled neuroimaging study can illustrate functional properties of a brain region whose necessity for a particular aspect of behaviour has been established with lesion data. There can arise, however, situations in which two different neuroimaging studies (and indeed, as we will review here, two different analyses of the same data set) produce mutually incompatible results, thereby leaving ambiguous the nature of the structure-function mapping that the experiments were designed to address. This is a situation in which the ability to locally alter brain function in a prospective manner can be particularly useful. The aspect of cognition that is at issue in this commentary is the short-term retention (STR, a.k.a. “storage” or “maintenance”) of information that is required by tests of short-term and working memory. One way to operationalise this construct is to vary the number of items that must be retained on different trials. Prompted by a review of studies that manipulated verbal memory load in this way, we applied two different analyses to the data from a sample of 24 subjects retaining 2 vs. 5 letters in a (7 sec) delayed-recognition task during functional magnetic resonance imaging (fMRI): a spatially normalized group-average (SNGA) analysis; and singlesubject (SS) analyses that treated each subject’s data as an individual case (Feredoes and Postle, 2007). The SNGA analysis revealed a region in left posterior middle frontal gyrus (MFG) of the prefrontal cortex (PFC), near the border of Brodmann Areas 9 and 6, that was reliably sensitive to the manipulation of load, and, therefore, a candidate locus for the STR of verbal information. (This also replicated the findings of previous studies that had also used SNGA analyses (e.g., Narayanan et al., 2005).) The SS analyses, in contrast, produced results that were topographically highly variable across subjects and, notably, did not include the left posterior MFG in any subject. Instead, the regions demonstrating load effects were best summarized as occurring in left posterior perisylvian cortex in the majority of subjects. These results prompted us to address the obvious question of which brain regions, those identified by the SNGA or the SS analyses, contributed more importantly to the STR of information in our task? (Although it was empirically plausible that the two sets of regions could make functionally comparable contributions, the two sets of results were difficult to reconcile theoretically (Postle, 2006).) To address this question, we performed a second study that would