On your data

Spacing model

CorrExplore begins by randomly placing each variable, then iteratively updates their positions using a variant of Multidimensional Scaling (MDS). This variant replaces the classic springs model used by most other MDS algorithms with a rubber bands and magnets model, a technique we adopt from research on co-citation network visualization (Noel, Chu, & Raghavan, 2002). See below for more technical details.

Without this modification, MDS is not a valid approach to mapping correlations. Despite their widespread use, the fundamental assumption made by MDS maps of correlation matrices is simply wrong: correlations are not distances, no matter what kind of arithmetic you use to transform them (typically 1-r, as if the only problem is that they can be negative). MDS doesn't just require positive numbers, it requires actual distances as inputs, and treating negative correlations as simply a greater distance than any positive correlation is a kludge that adds arbitrary rigidity to the model.

The solution: rubber bands and magnets

In the classic MDS spring model, each tie is modeled as a spring with its resting distance equal to the distance provided by the input matrix. As such, it will either push its endpoint variables apart or pull them together to reach that distance. CorrExplore's spacing model treats positive correlations as rubber bands, and negative correlations as repelling magnets. In other words, the repelling force of a negative correlation decreases as the two variables move apart, but unlike with a spring, no matter how far the variables move apart it does not become an attraction force. Similarly, positive correlations always act as attraction forces, no matter how close the variables become. What stops variables from collapsing on one another is the balance of forces in the network, not any arbitrary ideal distance.

Because of this reliance on a balance of forces, some proprietary code is necessary to detect special cases that would cause positioning problems, resulting in automatic adjustment of settings and even certain variables being removed from the model (usually one or two at most; right-click to see which ones). You can usually get the model to accept such a variable by adding in a few other variables that are positively correlated with it or raising the significance cutoffs.

Visually explore correlation networks