You are viewing the SNL 2017 Archive Website. For the latest information, see the Current Website.

Poster A51, Wednesday, November 8, 10:30 – 11:45 am, Harborview and Loch Raven Ballrooms

Test-retest reliability comparison of RSA and GLM approaches in a language task

Ryan Staples1, Einar Mencl1,3, Jeffery Malins1, Daniel Brennan1, Ken Pugh1,3,4, Robin Morris2;1Haskins Laboratories, 2Georgia State University, 3Yale University, 4University of Connecticut

We examined the rest-retest reliability of representational similarity analysis (RSA) and a standard general linear model (GLM) approach in parallel, as applied to a language task. The purpose was to examine the differences between the GLM’s smoothed, large-grained representation of neural information and RSA’s more fine-grained, pattern based representation. The subjects are 14 adults who were scanned twice, approximately 3 weeks apart, on the same tasks and who had no reading or speech disabilities. Subjects were presented with a picture and then with either a written or recorded word that did or did not match the picture, though only the mismatch conditions were considered in our analyses. The GLM reliability was calculated as the voxel-wise correlation between time one and time two print-speech contrast maps. RSA reliability was calculated as the voxel-wise correlation between time one and time two RSA maps, which themselves were created by identifying, on a trial-wise basis and using a searchlight analysis, which voxels had a similar response within the auditory and visual conditions, and a different response in the opposite condition. Both approaches demonstrated broad reproducibility across perisylvian areas, with some differences appearing. Both analyses showed a great deal of reliability in bilateral fusiform, lingual, and middle occipital gyri. The GLM showed broader areas of bilateral middle and superior temporal gyri reliability, whereas the RSA was focused in the right middle temporal gyrus. Regions of the inferior parietal lobule showed strong reliability in both analyses, though the precise location differed. We also calculated the voxel-wise correlation between time one GLM results and time one RSA results and the same between time two GLM and RSA results. We again found broad positive correlation in the right and left fusiform, lingual, and middle occipital gyri, as well as in the left inferior frontal gyrus. Notably, no strong correlation was found in temporal regions. These results suggest that both analysis methods are highly reliable for language tasks. Additionally, both types of analysis are reliably correlated with each other in language areas. The RSA results additionally suggest that the response in visual areas to written words is highly stereotyped and differentiable from those same regions’ response to auditory stimuli, but that the precise, local pattern of each response to an auditory word in temporal regions is more variable, more similar to those regions’ responses to visual language stimuli, or both.

Topic Area: Methods

Back to Poster Schedule