Abstract
In representational similarity analysis (RSA), participants are presented with multiple stimuli while brain responses to those conditions, or activity patterns, are measured. One of many available distance measures is applied to all pairs of activity patterns and organized in a representational dissimilarity matrix, or RDM. In most applications activity patterns are not direct measures of neural activity, but influenced by a measurement model, the process that transforms neural activity into measured activity patterns. This transformation distorts the true underlying representational geometry, and it would be useful to obtain information about what distance measures are most robust to such distortions, having better construct validity. We used simulation work to explore this issue. A biologically-motivated encoding model of face shape was presented with a database of realistic faces to obtain the “true” neural activity patterns, and a linear measurement model was used to obtain transformed activity patterns. In every simulation, we randomly sampled a different measurement model and computed RDMs in the neural and measurement spaces, using multiple available distances. We then computed the Spearman correlation between the two RDMs as a measure of construct validity. The highest values were obtained by the inner product, followed by Mahalanobis, city-block, and euclidean measures. Other measures produced relatively poor validity values. We determined the influence of multiple features of the measurement model on construct validity, including measurement noise, sparseness (number of zero weights per voxel), and number of informative and noninformative voxels. Construct validity was mostly invariant to changes in sparseness and number of informative and noninformative voxels, whereas it dropped with increments in measurement noise.