Perception of color emotions for single colors in red-green defective observers

Subjects

The participants, all Japanese, included 13 men and 10 women with normal vision, and 13 men who were red-green defective observers, including both the protan and deutan types. Age is a potential factor influencing color preference, especially between younger and older women (Hurlbert & Owen, 2003). Thus, participants were sampled within a range of 20–40 years as much as possible. The mean ages in the normal men, normal women, and red-green defective observers were 23.23 years (SD = 4.00), 26.50 years (SD = 6.70), and 23.92 years (SD = 7.01), respectively, within an age range of 18–43 years. All participants were tested using Ishihara pseudoisochromatic plates. In addition, red-green defective individuals underwent testing with the Anomaloscope (NEITZ OT-ll) and then were diagnosed with the type and severity of color vision deficiency according to their matching, identifying two protanopes, one deuteranope, and 10 deuteranomalous trichromats. Red-green defective participants included all types of defects and severity as listed above. Written informed consent was obtained from all participants on an experimental protocol that was approved by the Ethics Committee of Kagawa University (26-002).

Equipment and stimuli

A computer (DELL Optiplex 9020) controlled test instrument and a calibrated monitor (EIZO ColorEdge CX270 with a resolution of 1,900 × 1,200 pixels), were used. Stimuli included 32 colors from the BCP (Palmer & Schloss, 2010), composed of saturated, light, muted, and dark samples of eight hues: red (R), orange (O), yellow (Y), chartreuse (H), green (G), cyan (C), blue (B), and purple (P). These eight hues were made up of the four Hering primaries (Stockman & Brainard, 2010), R, G, B, and Y, as well as four well-balanced binary hues, O, H, C, and P. The four color sets (i.e., saturated, light, muted, and dark) were defined by cutting color space that differed in the saturation and lightness levels. The 32 color stimuli in the current experiment were simulated as closely as possible on the monitor, which was white calibrated as a reference (x = 0.31, y = 0.32, Y = 99.93 cd/m²). The chromaticity coordinates for the 32 color stimuli were measured with a luminance and color meter (KONICA MINOLTA CS-150) and are shown in Table S1. The experiment was carried out in the dark room of a laboratory, on a calibrated monitor, under D65 ceiling lighting with a vertical luminance on the display of approximately 350 lux. The color stimuli were randomly displayed as circular patches at a viewing distance of 50 cm (19.68 inches) and a visual angle of four degrees. A chin rest was used to fix the distance between the participant and the screen, and color stimuli were presented at the center of the monitor at the participant’s eye level.

Procedure

The participants completed two tasks, which included a color naming and an emotion rating task involving six subtasks for six emotions, with roughly 1-minute intervals between tasks. The experiment was replicated at least two hours after the first measurement, with a reversal of the task order. Color stimuli were presented individually on a gray background (x = 0.32, y = 0.32, Y = 19.07 cd/m²).

In the color naming task, participants selected one color name from the 11 BCTs using a computer mouse: red, yellow, green, blue, purple, brown, orange, pink, black, white, and gray. These BCTs have been used for color naming across different languages and geographical locations (Berlin & Kay, 1969; Uchikawa & Boynton, 1987), and even across different color vision types (Bonnardel, 2006; Lillo et al., 2001; Lillo et al., 2014). In the present experiment, the buttons for the 11 BCTs were presented in Japanese below the color stimuli: aka (red), midori (green), ki (yellow), ao (blue), daidai (orange), momo (pink), murasaki (purple), tya (brown), kuro (black), hai (gray), and shiro (white). Participants responded by clicking on the color name as rapidly as possible. The buttons representing the color names (with each size being 2.1 degrees horizontal and 1.8 degrees vertical) were presented in a horizontal line with the order mentioned above. The order and position of each button was not changed throughout the task.

During the emotion rating task, participants rated six emotions using the bipolar adjectives. Six adjective pairs were chosen from a previous study by Ou et al. (2004), and presented in Japanese as follows: kirei-kitanai (clean–dirty), shinsen-furukusai (fresh–stale), yawarakai-katai (soft–hard), suki-kirai (like–dislike), tumetai-atatakai (cool–warm), and karui-omoi (light–heavy). The rating task for each emotion was carried out as a separate subtask, and counterbalanced across participants. Participants were instructed to rate their emotion for each stimulus by moving the mouse cursor along a horizontal bar representing an emotional range from zero to ten (17.9 degrees horizontal and 1.6 degrees vertical). For instance, in the case of a suki-kirai (like–dislike) rating, zero and ten describe “totemo suki (like very much)” and “totemo kirai (dislike very much),” respectively. The default position was set to the center of the bar, indicating neutral.

In both tasks, before the color stimuli appeared, a 500-ms fixation cross (1.3 degrees diameter) was presented. In the color naming task, color stimuli were presented for 10-s. If there was no response within the 10-s duration, the next stimulus was presented. In the emotion rating task, participants performed without a time limit, but were asked to respond as quickly as possible.

Reliability and correlation between six color emotion ratings

Initially, in order to confirm the reliability of each color emotion rating, correlation coefficients were calculated between the first and second ratings for the six color emotions within each group (see reliability in Table 1). Reliability for all emotion ratings was high and significant in all groups (all p < .001). In addition, a correlation matrix was calculated to quantify the relationships between the ratings of the six adjectives for all groups (see correlation matrix between six emotions in Table 1). In the trichromatic male and red-green defective groups, only the warmth ratings showed no significant correlation to any other emotions. On the other hand, in the female observers, the hardness ratings also did not reveal any significant correlations to the other emotions except for the weight ratings. The other correlations were all significant (all p < .01).

Color emotion ratings for normal and red-green defective observers

Figures 1 and 2 show the average color emotion ratings within each group of the four sets of colors (i.e., saturated, light, muted, and dark) along all eight hues (i.e., R, O, Y, H, G, C, B, and P). For cleanliness (clean–dirty) ratings in the trichromatic individuals, including men and women, the saturated and light samples were assessed as cleaner, with the exception of saturated purple, compared with the muted and dark colors, which were assessed as dirtier, regardless of hue. For both men and women, dark orange was rated the dirtiest. The red-green defective observers’ ratings were more varied across the four color sets, specifically in the green to cyan range, than were those of the normal participants. For the red-green defects, dark red was identified as dirtier, while light blue was the cleanest.

In the freshness (fresh–stale) ratings for the female group, there was a clear difference between each of the four color sets in the range of red to yellow. Similar to the cleanliness ratings, normal individuals rated the saturated and light samples as fresher, with the exception of saturated purple, while the muted and dark samples were rated as staler. For the red-green defective observers, light blue was rated the freshest, which was the same as in the cleanliness ratings, while dark orange was the stalest.

As for hardness (soft–hard), the normal men and red-green defective observers rated the light color sets as all softer, while the dark color sets were rated as harder. On the other hand, the female observers rated the saturated and dark color samples as harder compared with the light and muted sets.

For the preference (like–dislike) ratings, in order to facilitate comparison with previous results, we converted the rating scores to the inverse scale (0 for like, 10 for dislike). Normal individuals liked the saturated and light color sets, with the exception of saturated purple. Conversely, the muted and dark sets, with hues from red to chartreuse, were disliked. For the red-green defective observers, the preference pattern did not fluctuate much along the hue axis. Saturated orange was the most liked, similar to the men with normal vision, and muted purple was the least liked.

For all groups, the warmth (cool–warm) rating patterns steadily decreased from warm to cool in the red to blue range, but purple increased in warmth. Male observers tended to rate red, orange, and yellow hues in decreasing warmth according to the following order: saturated (warmest), light, muted, and dark (coolest). In addition, the women’s ratings varied across four sets in green and cyan hues. The four color sets (i.e., saturated, light, muted, and dark) were not as clearly separated by warmth ratings in the red-green defective observers.

For weight (light–heavy) ratings, all groups seemed to assign weight based on the four color sets. The light colors were rated lighter, while the dark color sets were heavier.

Mixed model ANOVAs were used to assess differences between groups in the six emotion ratings. For each emotion rating, a mixed-model ANOVA with the four color sets (i.e., saturated, light, muted, and dark), eight hues (i.e., R, O, Y, H, G, C, B, P), and three groups (male, female, and red-green defective observers) was conducted. Since the primary aim of this study was to examine differences between groups in color emotion ratings, only significant interaction effects that involved group differences (i.e., color set × group, hues × group, or color set × hues × group) were evaluated and are discussed here. Therefore, main effects of color set and hues, as well as the color set × hue interaction are not discussed. In addition, observed differences between groups based on Bonferroni-corrected pairwise comparisons are shown in Table 2.

For cleanliness ratings, the analysis showed a significant three-way interaction between group, color set, and hue ($F\left(42,693\right)=2.17,p<.001,{\eta }_G^2=0.04$). To examine the three-way interaction, post-hoc two-way ANOVAs using the mean ratings of each hue were conducted with set and group as factors. The results revealed a significant interaction of group and color set on orange, yellow, chartreuse, green, and cyan (orange: $F\left(4.71,77.74\right)=5.69,p<.001,{\eta }_G^2=0.18$; yellow: $F\left(3.70,61.11\right)=3.46,p<.05,{\eta }_G^2=0.12$; chartreuse: $F\left(4.33,71.47\right)=3.85,p<.01,{\eta }_G^2=0.12$; green: $F\left(4.38,72.23\right)=2.99,p<.05,{\eta }_G^2=0.10$; cyan: $F\left(4.67,77.05\right)=7.19,p<.001,{\eta }_G^2=0.16$). Simple effects showed that there were group differences on the dark sets for orange, yellow, chartreuse, green, and cyan (orange: $F\left(2,33\right)=10.09,p<.001,{\eta }_G^2=0.38$; yellow: $F\left(2,33\right)=6.67,p<.01,{\eta }_G^2=0.29$; chartreuse: $F\left(2,33\right)=11.25,p<.001,{\eta }_G^2=0.41$; green: $F\left(2,33\right)=3.90,p<.05,{\eta }_G^2=0.19$; cyan: $F\left(2,33\right)=3.73,p<.05,{\eta }_G^2=0.18$), on the muted sets for yellow, and chartreuse (yellow: $F\left(2,33\right)=3.97,p<.05,{\eta }_G^2=0.20$; chartreuse: $F\left(2,33\right)=9.18,p<.001,{\eta }_G^2=0.36$), and on the saturated sets for cyan ($F\left(2,33\right)=8.56,p<.01,{\eta }_G^2=0.34$).

For freshness ratings, the interaction of group and color set was significant ($F\left(6,99\right)=9.93,p<.001,{\eta }_G^2=0.34$). Post-hoc one-way ANOVAs with group as a factor were done on the mean ratings of the four color sets. These analyses showed that there were significant group differences on the saturated, muted, and dark sets (saturated: $F\left(2,33\right)=13.12,p<.001,{\eta }_G^2=0.16$; muted: $F\left(2,33\right)=6.94,p<.01,{\eta }_G^2=0.08$; dark: $F\left(2,33\right)=7.24,p<.01,{\eta }_G^2=0.14$).

The ANOVA for the hardness ratings revealed the significant interaction between group, color set, and hue ($F\left(42,693\right)=1.75,p<.01,{\eta }_G^2=0.03$). To examine the three-way interaction, post-hoc two-way ANOVAs using the mean ratings of each hue were conducted with set and group as factors. The results revealed a significant interaction of group and color set on red, orange, yellow, chartreuse, and cyan (red: $F\left(6,99\right)=6.84,p<.001,{\eta }_G^2=0.19$; orange: $F\left(4.64,76.51\right)=4.30,p<.01,{\eta }_G^2=0.15$; yellow: $F\left(3.95,65.24\right)=3.71,p<.01,{\eta }_G^2=0.14$; chartreuse: $F\left(4.79,79.10\right)=7.56,p<.001,{\eta }_G^2=0.23$; cyan: $F\left(4.24,69.98\right)=3.10,p<.05,{\eta }_G^2=0.08$). Simple effects showed that there were group differences on the saturated color sets for red, orange, yellow, chartreuse, and cyan (red: $F\left(2,33\right)=9.20,p<.001,{\eta }_G^2=0.36$; orange: $F\left(2,33\right)=4.63,p<.05,{\eta }_G^2=0.21$; yellow: $F\left(2,33\right)=3.84,p<.05,{\eta }_G^2=0.19$; chartreuse: $F\left(2,33\right)=19.38,p<.001,{\eta }_G^2=0.54$; cyan: $F\left(2,33\right)=4.24,p<.05,{\eta }_G^2=0.20$), on the light sets for red and orange (red: $F\left(2,33\right)=5.30,p<.05,{\eta }_G^2=0.24$; orange: $F\left(2,33\right)=4.04,p<.05,{\eta }_G^2=0.20$) on the muted sets for red ($F\left(2,33\right)=5.47,p<.01,{\eta }_G^2=0.25$), and on the dark sets for yellow and chartreuse (yellow: $F\left(2,33\right)=4.00,p<.05,{\eta }_G^2=0.20$; chartreuse: $F\left(2,33\right)=6.75,p<.01,{\eta }_G^2=0.29$).

For the preference (like-dislike) ratings, there was a significant three-way interaction between group, color set, and hue ($F\left(42,693\right)=1.70,p<.01,{\eta }_G^2=0.03$). The post-hoc two-way ANOVAs using the mean ratings of the eight hues were conducted with color sets and group as factors. These results showed that for the cyan color, there was a significant group difference ($F\left(2,33\right)=5.30,p<.05,{\eta }_G^2=0.12$) and for the purple color, group and color set interacted significantly ($F\left(6,99\right)=2.54,p<.05,{\eta }_G^2=0.08$). Simple effects revealed group differences on the saturated color only ($F\left(2,33\right)=5.17,p<.05,{\eta }_G^2=0.24$).

The ANOVAs for the warmth (cool-warm) ratings also revealed a significant three-way interaction between group, color sets and hue ($F\left(42,693\right)=1.44,p<.05,{\eta }_G^2=0.03$). The post-hoc two-way ANOVAs with color sets and group as factors were done on the mean ratings of the eight hues. These results revealed that for the cyan color, group and color set interacted ($F\left(4.07,67.20\right)=2.29,p=.067,{\eta }_G^2=0.07$). Simple effects showed group differences on the saturated set ($F\left(2,33\right)=5.57,p<.01,{\eta }_G^2=0.25$). Furthermore, there was a significant group difference for purple ($F\left(2,33\right)=3.57,p<.05,{\eta }_G^2=0.04$).

For weight ratings, the interaction between group and color set was significant ($F\left(4.42,72.96\right)=5.34,p<.001,{\eta }_G^2=0.08$). Post-hoc one-way ANOVAs showed that there was a significant group difference for the light and dark sets (light: $F\left(2,33\right)=5.96,p<.01,{\eta }_G^2=0.14$; dark: $F\left(2,33\right)=10.44,p<.001,{\eta }_G^2=0.18$).

Prediction of color emotions based on three cone-contrast values

Similar to previous studies (Hurlbert & Ling, 2007; Palmer & Schloss, 2010; Taylor, Clifford & Franklin, 2013; Álvaro et al., 2015), the six emotion ratings were examined in relation to the three cone-contrast mechanisms based on a trichromatic vision system. This mechanism represents pairs of mutually exclusive perceptual categories of red-green (L-M), blue-yellow (S-(L+M)), and luminance cone-contrasts (L+M). Hurlbert & Ling (2007) summarized color preference by weights on two cone-contrasts, and Álvaro et al. (2015) and Taylor, Clifford & Franklin (2013) examined color preferences in terms of two cone-contrasts, lightness, and saturation values. Palmer & Schloss (2010) also used the two cone-contrast values, lightness, and saturation as predictors. In the present study, the two dimensions consisting of L-M and S-(L+M) cone-contrasts and the luminance cone-contrast value (lightness) was included. Because color emotions such as hardness or weight ratings may have a strong relation to lightness (Ou et al., 2004; Gao & Xin, 2006), and lightness of color has an influence on most color emotions (Xin et al., 2004).

Three cone-contrast values were calculated as follows according to the procedure of Álvaro et al. (2015). Firstly, the CIE x, y, and Y values were translated to the amount of L, M, and S cone excitations using Smith and Pokorny’s cone fundamentals (Smith & Pokorny, 1975). Secondly ΔL, ΔM, and ΔS, representing differences between each stimulus and the background were calculated using the following equations: ΔL = (L_s − L_b)∕L_b, ΔM = (M_s − M_b)∕M_b, and ΔS = (S_s − S_b)∕S_b. The subscript ‘s’ in the equations stands for the stimulus, while the subscript ‘b’ stands for the background color. Lastly, three cone-contrast values were calculated according to the cone-contrast weights that had been defined by Eskew, Mclellan & Giulianini (1999). Figure 3 shows the three cone-contrast values for the four color sets and eight hues.

The three cone-contrast values were used as predictors of the mean ratings for the six color emotions averaged within each group using linear regression. The resultant standardized regression coefficients, coefficients of determination (R squared), and F statistics are shown in Fig. 4. The prediction model that included the three cone-contrasts accounted for all emotion ratings significantly. Specifically, for the warmth and weight ratings, this model accounted for the high variances by more than 87% in all groups. The models for the cleanliness and hardness ratings also accounted for the high variances within a range of 65–85% with the exception of the hardness rating in the female group. For the freshness ratings, the model explained just half of the variances in the trichromatic group, while a higher variance was found in the red-green defective observers. Lastly, for the preference ratings, variances were lower, with the model accounting for about 40% in trichromatic observers, and 55% in the red-green defective group.

In all of the regression models, except for the warmth ratings, the L+M contrast was a significant predictor of color emotion for all groups. The cleanliness model indicated that the S-(L+M) and L+M cone-contrasts were significant predictors of cleanliness ratings in all groups. In the freshness model, the S-(L+M) cone-contrast was a significant predictor of the freshness ratings in the trichromatic male and red-green defective group, but not in the female group. In the hardness model, the S-(L+M) cone-contrast was significant in the male and red-green defective groups. For the trichromatic male observers, the L-M cone-contrast was also a significant predictor. In the preference model, the L-M and S-(L+M) cone-contrasts were not significant predictors of preference ratings in any group. In the weight model, the S-(L+M) cone-contrast was a significant predictor of weight ratings for all groups. Lastly, for the warmth model in all groups, the L-M and S-(L+M) cone-contrasts were significant predictors of warmth ratings, but L+M was not significant. The regression weight of the red-green defective observers was much lower than those of trichromatic men or women.

Color naming results and the relation to color emotion ratings

Table 3 provides the color naming results of the normal including men and women and red-green defective groups. Red-green defective group included all types of defects and severity, following the evidence that protanope and deuteranope groups have similar color naming, with such similarities also apparent between protanomalous and deuteranomalous (Nagy et al., 2014). The naming responses and their frequencies are described in the table. If the naming responses of the red-green defective participants were not seen in the group with normal vision, they were underlined in the table. There were no missing response data in all participants.

To examine group differences in naming responses for the 32 colors, Fisher’s exact tests were conducted (2 groups × 11 BCTs). These results showed that naming responses were significantly different between groups for the light orange (p < .05), the dark yellow (p < .05), the saturated cyan (p < .001), the light cyan (p < .001), and the muted cyan (p < .001) colors.

Additionally, several variables were examined within each group in relation to performance on color naming, including response time, naming consistency, and naming consensus for each color stimulus, similar to what was done in the study by Álvaro et al. (2015). Naming consistency refers to the number of participants that used the same color name in the first and second responses. Naming consensus is the number of participants within a group that agreed with the most frequently used response. Correlation coefficients among these variables were calculated. There was a significant correlation between response time and consensus in all groups (men: r =  − 0.79, p < .001; women: r =  − 0.65, p < .001; red-green defects: r =  − 0.74, p < .001). There was no significant correlation between response time and consistency in all groups (men: r = 0.11, n.s.; women: r =  − 0.06, n.s.; red-green defects: r =  − 0.12, n.s.). Also, there was no significant correlation between consistency and consensus in all groups (men: r = 0.18, n.s.; women: r = 0.25, n.s.; red-green defects: r = 0.33, n.s.). Finally, correlation coefficients were calculated between the variables that influenced color naming performance and each emotion rating. As shown in Table 4, there were only significant correlations between the hardness ratings and response time, as well as between the hardness ratings and consensus, in the female group only.