Developmental Divergence in Univariate and Multivariate fMRI Analyses

A 2010 FINS paper from Cohen et al. demonstrates that multivariate patterns in neural recruitment during response inhibition across the brain are significantly predictive of response inhibition ability and age of the scanned subject, and shows that other factors (such as response variability and reaction times) cannot be similarly predicted from the same data.

Cohen et al asked twenty-seven 9-19 year-olds and nine 25-30 year-olds to complete a standard "stop signal" task. On 75% of trials ("Go trials") this task is identical to a 2-choice reaction time task (i.e., as quickly as possible press button 1 for stimulus A and button 2 for stimulus B) but on the remainder of trials ("Stop trials"), these stimuli are followed by a tone with a variable interval. This tone indicates that subjects must attempt to stop all responses on that trial; if subjects fail to stop, the tone is presented 50ms earlier the next time it occurs, whereas if they successfully stop the tone is presented 50ms later. This adaptive algorithm will eventually converge on a stop signal delay that yields 50% successful stops, which can then be integrated with the reaction times on the other 75% of trials to arrive at "Stop Signal Reaction Time" (SSRT), a well-validated measure of how long it takes subjects in order to stop a response.

Estimation of SSRT requires some tricky assumptions, and it may be influenced by the variability of responding on trials without the stop signal. Furthermore, it's unclear whether age-related improvements in SSRT might occur via changes in response variability, Go trial reaction time, or be truly specific to changes in the efficiency of stopping-related processes.

fMRI analyses revealed about what you'd expect: SSRT and age were not correlated with the neural recruitment observed on Go trials, which primarily activated motoric regions of the frontal lobe. On the other hand, longer Go reaction times were related to less activity in the right posterior middle frontal gyrus (going against the time-on-task concerns often raised regarding neuroimaging results). Increased response time variability was also related to less left parietal activation. Multivariate analyses revealed no additional information about the neural substrates of Go reaction times or response variability.

In contrast, massive swaths of cortex were recruited on successful Stop trials, including virtually the entire right frontal lobe. Here's a glimpse of how much cortex we're talking about:

i-978cd711774fd692f7b21484a2672baf-Cohen.jpg

What's really impressive here is the discrepancy between the way that SSRT and age relate to this pattern, depending on whether one uses univariate or multivariate analyses. At the univariate level, age was negatively correlated with recruitment of a single region in the left rostral anterior cingulate; in anterior parts of the brain, SSRT was negatively correlated primarily with medial areas, both frontal and striatal. So are we to believe that lateral frontal cortex, the foremost substrate for executive function, is unrelated or only weakly related to factors as important as age and response inhibition? Not so fast. When you look at multivariate relationships of age & SSRT to neural activity, this is what you see:

i-55f4ceea1f0eacd6cf609ab410f07b3e-Cohen1.jpg

OK, so now we're to believe that the entire lateral frontal cortex is related to individual differences in age and response inhibition!

Within the field of developmental neuroscience and psychology, this study is good, incremental progress: Cohen et al have shown that neural activity during response inhibition predicts age and SSRT roughly equally well, but that response variability and Go reaction times are not likely the whole story.

The results pose a more substantive challenge to the larger field of neuroimaging, in which we're going to have to grapple with the fact the relationships of subject characteristics to neural activity seems so dependent on the statistical methods we use. It will be difficult to understand a convergent picture of brain function when the results from univariate and multivariate methods seems so divergent.

More like this