Studying the Separability of Visual Channel Pairs in Symbol Maps

TL;DR

A crowdsourced experiment is presented that evaluates the separability of four channel pairs--color (ordered) x shape, color (ordered) x size, size x shape, and size x orientation--in the context of bivariate symbol maps, finding that color x shape is the most separable and size x orientation the least separable.

Abstract

Visualizations often encode multivariate data by mapping attributes to distinct visual channels such as color, size, or shape. The effectiveness of these encodings depends on separability--the extent to which channels can be perceived independently. Yet systematic evidence for separability, especially in map-based contexts, is lacking. We present a crowdsourced experiment that evaluates the separability of four channel pairs--color (ordered) x shape, color (ordered) x size, size x shape, and size x orientation--in the context of bivariate symbol maps. Both accuracy and speed analyses show that color x shape is the most separable and size x orientation the least separable, while size x color and size x shape do not differ. Separability also proved asymmetric--performance depended on which channel encoded the task-relevant variable, with color and shape outperforming size, and square shape especially difficult to discriminate. Our findings advance the empirical understanding of visual separability, with implications for multivariate map design.

Figures (4)

• Figure 1: Mean accuracy with 95% CIs (left) and mean log response time with 95% CIs (right) for the four visual channel pairs. Significant pairwise contrasts (Tukey-adjusted) are annotated. Note: Both the accuracy axis and the log response time axis do not start at zero. Both axes are shown over the observed range to highlight meaningful variation in the data.
• Figure 2: Mean accuracy with 95% CIs for each channel pair and flip, broken down by the value of the task-irrelevant (Dimension B) attribute carried by the target symbol. Each panel represents one channel pair; within each panel, the three x-axis positions correspond to low, medium, and high levels of Dimension B. Colored points represent flips (i.e., which channel encodes the task-relevant Dimension A). To aid interpretation, the symbol beneath each point shows the actual Dimension B value paired with the target symbol in that condition. Higher accuracy indicates that the Dimension A value was easier to recognize when paired with that particular Dimension B value. Note: The accuracy axis does not start at zero and is shown over the observed range to highlight meaningful variation in the data.
• Figure 3: Mean log response time with 95% CIs for each channel pair and flip, broken down by the value of the task-irrelevant (Dimension B) attribute carried by the target symbol. Panels correspond to channel pairs, and colored points indicate the flip (task-relevant channel). As in the accuracy plots, the symbol directly beneath each mean point depicts the specific Dimension B value associated with that target symbol. Lower values on the y-axis indicate faster responses. Note: The log response time axis does not start at zero and is shown over the observed range to highlight meaningful variation in the data.
• Figure 4: Within-flip effects of Dimension B (task-irrelevant channel) on recognition of the target value of Dimension A, for each channel pair. Each panel shows the mean accuracy and log response time differences between the low, medium, and high Dimension B values within each flip. Points represent contrast estimates (low–medium, low–high, medium–high), with horizontal bars showing 95% CIs. The symbols in the legend to the right indicate the specific Dimension B value levels compared in each contrast. Positive $\Delta$Accuracy values indicate higher accuracy for the first value in the pair (e.g., low $>$ high), while negative values indicate higher accuracy for the second value. For $\Delta$Log response time, positive values indicate slower responses for the first value in the pair and faster responses for the second; negative values indicate the opposite pattern. Stars denote Tukey-adjusted significance: $* = p < .05$, $** = p < .01$, $*** = p < .001$.