The relation between color and spatial structure for interpreting colormap data visualizations

Abstract

Interpreting colormap visualizations requires determining how dimensions of color in visualizations map onto quantities in data. People have color-based biases that influence their interpretations of colormaps, such as a dark-is-more bias-darker colors map to larger quantities. Previous studies of color-based biases focused on colormaps with weak data spatial structure, but color-based biases may not generalize to colormaps with strong data spatial structure, like "hotspots" typically found in weather maps and neuroimaging brain maps. There may be a hotspot-is-more bias to infer that colors within hotspots represent larger quantities, which may override the dark-is-more bias. We tested this possibility in four experiments. Participants saw colormaps with hotspots and a legend that specified the color-quantity mapping. Their task was to indicate which side of the colormap depicted larger quantities (left/right). We varied whether the legend specified dark-more mapping or light-more mapping across trials and operationalized a dark-is-more bias as faster response time (RT) when the legend specified dark-more mapping. Experiment 1 demonstrated robust evidence for the dark-is-more bias, without evidence for a hotspot-is-more bias. Experiments 2 to 4 suggest that a hotspot-is-more bias becomes relevant when hotspots are a statistically reliable cue to "more" (i.e., the locus of larger quantities) and when hotspots are more perceptually pronounced. Yet, comparing conditions in which the hotspots were "more," RTs were always faster for dark hotspots than light hotspots. Thus, in the presence of strong spatial cues to the locus of larger quantities, color-based biases still influenced interpretations of colormap data visualizations.

Figure 1.

(A) Example trial from Schloss et al. (2019). Participants saw colormaps with legends and indicated whether there were more alien animal sightings early or late in the day. (B) Example trial in the present study. Participants saw colormaps with legends and indicated whether there were more alien animal sightings on the left or right side of the map.

Figure 2.

Predicted RTs for dark-more encoding (D+, black bars) and light-more encoding (L+, white bars) for dark and light hotspots (x-axis), depending on if there is (A) only a dark-is-more bias, (B) only a hotspot-is-more bias, or (C) a combination of both a dark-is-more and hotspot-is-more bias. Conditions in which the colors in the hotspot map to larger quantities are indicated by bold text (i.e., D+ when the hotspots are dark and L+ when hotspots are light).

Figure 3.

Illustration of the process used to generate colormaps with hotspot configurations and scrambled configurations (see the description in the text for details).

Figure 4.

The top row shows the three color scales tested in this study: Autumn (Experiments 1–2), Hot (Experiments 3–4), and Viridis (Experiments 1–4). These color scales are applied to the same underlying data sets to produce colormaps with dark hotspots (middle row) and light hotspots (bottom row). The three left columns show colormaps constructed from underlying data sets with the initial noise level used in Experiments 1 to 3, and the two right columns show colormaps constructed from underlying data sets with reduced noise used in Experiment 4.

Figure 5.

Mean RTs for the (A) hotspot configurations and (B) scrambled configurations from Experiment 1. Black bars represent dark-more encoding (D+) and white bars represent light-more encoding (L+). Bars are grouped according to hotspot lightness (or scrambled hotspot lightness) (x-axis). Data are separated by the Autumn and Viridis color scales. Conditions in which the colors in the hotspot map to larger quantities are indicated by bold text (i.e., D+ when the hotspots are dark and L+ when hotspots are light). Error bars represent standard errors of the means using the Cousineau (2005) adjustment to account for overall differences in RT at the subject level.

Figure 6.

Mean RTs from Experiment 2 for (A) balanced cue images and (B) hotspot-more images, separated for dark-more encoding (D+, black bars) and light-more encoding (L+, white bars) depending on whether the hotspots were dark or light (x-axis). Data are separated by the Autumn and Viridis color scales. Conditions in which the colors in the hotspot map to larger quantities are indicated by bold text (i.e., D+ when the hotspots are dark and L+ when hotspots are light). Error bars represent standard errors of the means using the Cousineau (2005) adjustment to account for overall differences in RT at the subject level.

Figure 7.

Mean RTs for the hotspot localization task in (A) Experiment 2, (B) Experiment 3, and (C) Experiment 4. Data are plotted separately for Autumn and Viridis in Experiment 2 and for Hot and Viridis in Experiments 3 and 4 for dark hotspots (dark gray bars) and light hotspots (light gray bars). Error bars represent standard errors of the means using the Cousineau (2005) adjustment to account for overall differences in RT at the subject level.

Figure 8.

Mean RTs from Experiment 3 for (A) balanced cue images and (B) hotspot-more images, plotted in the same manner as in Figure 6.

Figure 9.

Mean RTs from Experiment 4 for (A) balanced cue images and (B) hotspot-more images, plotted in the same manner as in Figures 6 and 8.