Simulating the Fidelity of Data for Large Stimulus Set Sizes and Variable Dimension Estimation in Multidimensional Scaling

Multidimensional scaling (MDS) is a statistical technique commonly used to model the psychological similarity among sets of stimulus items. Typically, MDS has been used with relatively small stimulus sets (30 items or fewer), in part due to the laborious nature of computational analysis and data collection. Modern computing power and newly advanced techniques for speeding data collection have made it possible to conduct MDS with many more stimuli. However, it is as yet unclear if MDS is as well-equipped to model the similarity of large stimulus sets as it is for more modest ones. Here, we conducted 337,500 simulation experiments, wherein hypothetical “true” MDS spaces were created, along with error-perturbed data from simulated “participants.” We examined the fidelity with which the spaces resulting from our “participants” captured the organization of the “true” spaces, as a function of item set size, amount of error in the data (i.e., noise), and dimensionality estimation. We found that although higher set sizes decrease model fit (i.e., they produce increased “stress”), they largely tended to increase determinacy of MDS spaces. These results are predicated, however, on the appropriate estimation of dimensionality of the MDS space. We argue that it is not only reasonable to adopt large stimulus set sizes but tends to be advantageous to do so. Applying MDS to larger sets is appealing, as it affords researchers greater flexibility in stimulus selection, more opportunity for exploration of their stimuli, and a higher likelihood that observed relationships are not due to stimulus-specific idiosyncrasies.