Looking on the Bright Side: Replicating the Association between Brightness and Positivity

We based our power analysis on the effects reported by Specker et al. (2018), the smallest effect size we reported in that study was a Cohen’s dz of .90, this was the effect size of the explicit association between brightness and positivity when using black and white stimuli in a Western cultural context. To err on the side of caution we used a slightly smaller estimated effect size of .80 for our power analysis using the G*Power software (Faul, Erdfelder, Lang, & Buchner, 2007). This indicated that we needed 15 participants to achieve a power of .80 with an error rate probability of 0.05. We therefore collected data from 15 participants. All participants were psychology students of the University of Vienna and received course credit for their participation. Eight participants were female and the mean age was 22.93 (SD = 3.96). The experiment was carried out in accordance with the Declaration of Helsinki and in a procedure which was approved by the local ethical committee of the University of Vienna. We pre-registered an exclusion criterion of a minimum accuracy of 60%, meaning that anyone below that threshold would be excluded. Since none of our participants met our exclusion criteria we included all participants in our analysis.

Our method, materials and procedure was identical to Study 1 of Specker et al. (2018), thus we used an IAT the only difference being that we included an explicit rating task. The IAT and the explicit rating task were counterbalanced. Meaning that half of the participants first completed the IAT and then gave explicit ratings and the other half first gave explicit ratings and then completed the IAT.

IAT task. We used the standard IAT design, that includes error feedback for incorrect responses. In this case, the error message remains on the screen until the correct key is pressed. Key assignments were visible throughout the task as recommended by Gawronski, Deutsch, and Banse (2011). We opted to counterbalance the key assignments in order to be able to eliminate block order as potential confound. Meaning that half of the participants first had a congruent blocks (i.e. bright/positive and dark/negative) and then incongruent blocks (i.e. bright/negative and dark/positive) and the other half first had incongruent blocks followed by a congruent blocks.

Explicit Rating Task. In the explicit rating task people were asked to rate each stimulus on positivity. People were asked: “How positive is this color patch?” and answered on a 7 point scale ranging from 1 (negative) to 7 (positive).

Stimuli. We created 18 color patches differing on brightness as stimuli for the IAT task. All color patches were blue. Nine color patches represented the bright category and nine the dark category. The bright category ranged from 10–30% brightness with increasing steps of about 2–3%. The dark category ranged from 70–90% brightness with increasing steps of about 2–3%. All images can be found here: https://osf.io/w8479/. As a manipulation check all stimuli were measured by use of a colorimeter (i1Xtreme by X-Rite) this was done by presenting the stimuli full screen on a CRT monitor and then taking the measurements. Measurement was carried out on the monitor that was used by the participant during the test. For this manipulation check we focused on the luminance measures of the LAB color model. This showed that the dark category was lower in luminance (and therefore darker) than the bright category, t(16) = –15.49, p < .001. All stimuli were presented against a grey background (RBG: 128, 128, 128; LAB: 64.3, –4.4, –23.1). To ensure consistency of color presentation across subjects the IAT task was conducted in a windowless room with constant lighting conditions and on a CRT Monitor.

To represent the positive and negative category, we selected words from the Berlin Affective Word List Reloaded (BAWL-R, Võ et al., 2009). The BAWL-R is a cross-validated resource of affective words in German that have been rated for valence and arousal. Negative words included words such as asocial, sad, and loveless (respectively, “asozial”, “traurig”, and “lieblos” in German) and the positive words included words such as loved, great, and coureagous (respectively, “beliebt”, “toll”, “mutig” in German). The words selected to be in the positive and negative category differed in valence, t(16) = 63.41, p < .001, but not in arousal, p = .77. We matched for word length, on average negative words had a length of 7.22 characters and positive words of 6.44 characters, this difference in length was not significant, p = .244.

Experimental scripts and related stimuli material is freely available on the Open Science Framework: https://osf.io/w8479/.¹

As in Specker et al. (2018) a D-score was calculated as recommended by Greenwald, Nosek, and Banaji (2003) for every person. The D-score ranges from –1.5 to +1.5. Typically, D-scores of > .65 are seen as large effects. Since the IAT is a relative measure a positive D- score indicates a larger association between brightness and positivity than brightness and negativity. As a consequence this also means a larger association between darkness and negativity than darkness and positivity. This indicated a large effect with the mean D-score in our sample being .65 (SD = .37). As in Specker et al. (2018) we then conducted a one-sided one sample t-test. This showed a significant difference from 0 in the expected (positive) direction, t(14) = 6.58, p < .001, Cohen’s dz = 1.73,² CI [.91, 2.51]. For the explicit association, as in Specker et al. (2018), a difference score was created by subtracting the standardized mean rating of the dark category from the standardized mean rating of the bright category. We then conducted a one-sided one sample t-test over the difference score. This did not show a significant difference from 0, t(14) = –.90, p > .05, Cohen’s dz = 0.23 CI [–.74, .28]. Finally, as in Specker et al. (2018) we computed the correlation between implicit and explicit association. The correlation between implicit and explicit association was r = –.39, p > .05, as illustrated in Figure 1.

Study 1 replicated the findings of Study 1 of Specker et al. (2018) meaning that both studies found a large effect for the implicit association of brightness and positivity. Study 1 did not show an explicit effect for the association between brightness and positivity that should theoretically have been there, we therefore interpret this as a failure to replicate. The most probable reason for this failure to replicate – besides a simple Type 2 error – is the fact that when varying brightness we did not perfectly control for variation in hue or saturation. Therefore, it is possible that effects found are due to either hue and/or saturation rather than brightness. This was also the main reason of the original study to perform a second study with achromatic stimuli. If similar effects are found with achromatic stimuli, one can be more secure in arguing that this was due to brightness. For an absence of the effect, it might be that because these stimuli were less controlled, the effect disappeared. This becomes more likely when combined with the findings of the original study that the explicit rating effects are smaller than the implicit rating effects. Though it has to be mentioned that the smaller effect size is likely due, at least in part, to the fact that reaction time based measures with a lot of trials have less variance and therefore larger effect sizes than a one-trial explicit rating. Another likely explanation for the absence of an explicit effect, is that in the explicit rating task people confounded “positivity” with “liking”. In other words, if someone saw an image they liked that was dark they may have perceived it as more positive, similarly, when they saw a bright image they disliked they may have perceived it as more negative. In any case, this seems to suggest that stimulus type and measurement type influence the size of the effect. In this case, the odds were stacked against finding the effect since it combined lesser controlled stimuli as well as explicit rating as a measurement. This is a notion we will investigate in more detail later by use of meta-analysis.

We based our power analysis on the effects reported by Specker et al. (2018), the smallest effect size we reported in that study was a Cohen’s dz of .90, this was the effect size of the explicit association between brightness and positivity when using black and white stimuli in a Western cultural context. To err on the side of caution we used a slightly smaller estimated effect size of .80 for our power analysis using the G*Power software (Faul et al., 2007). This indicated that we needed 15 participants to achieve a power of .80 with an error rate probability of 0.05. We therefore collected data from 15 participants. All participants were psychology students of the University of Vienna and received course credit for their participation. Eleven participants were female and the mean age was 19.67 (SD = 1.39). The experiment was carried out in accordance with the Declaration of Helsinki and in a procedure which was approved by the local ethical committee of the University of Vienna. We pre-registered an exclusion criterion of a minimum accuracy of 60%, meaning that anyone below that threshold would be excluded. Since none of our participants met our exclusion criteria we included all participants in our analysis.

Our method, materials and procedure was identical to Study 2 of Specker et al. (2018) and Study 1 of the current paper with the exception of the color stimuli. Rather than having variations of blue we constructed achromatic stimuli. These employed a ratio from black-to-white. The bright category consisted of 8 color patches which were filled 10 or 20% black (with 90 or 80% white respectively). The dark category consisted of 8 color patches which were filled 10 or 20% white (with 90 or 80% black respectively). All images can be found here: https://osf.io/w8479/. Again, the IAT task and the explicit rating task were conducted in a windowless room with constant lighting conditions on a CRT monitor.

We used the same analyses as in Specker et al. (2018) and Study 1. Thus, we calculated a D-score for every person as recommended by Greenwald, Nosek, and Banaji (2003). This indicated a large effect with the mean D-score in our sample being .95 (SD = .23). We then conducted a one-sided one sample t-test. This showed a significant difference from 0 in the expected (positive) direction, t(14) = 15.77, p < .001, Cohen’s dz = 4.07, CI [2.49, 5.64]. For the explicit association, we created a difference score by subtracting the standardized mean rating of the dark category from the standardized mean rating of the bright category. We then conducted a one-sided one sample t-test over the difference score. This showed a significant difference from 0 in the expected (positive) direction, t(14) = 11.88, p < .001, Cohen’s dz = 3.06, CI [1.83, 4.29]. Finally, we computed the correlation between implicit and explicit association. The correlation between implicit and explicit association was r = .63, p < .01, as illustrated in Figure 2. As can be seen from Figure 2 this correlation was mainly driven by one data point (the one on the far left), to test this hypothesis exploratively we calculated the correlation without this outlier. Without the outlier the correlation was r = .067.

Study 2 replicated the findings of Study 2 of Specker et al. (2018) meaning that both studies found a large effect for the implicit and explicit association of brightness and positivity. Taking these findings together with the findings of Study 1 we can conclude that we replicated the effect in 3 out of 4 cases. There are two possible explanations as to why we were able to replicate the explicit rating effects in Study 2, one would be that it is due to the higher level of control over our stimuli, the other option would be that positivity (i.e. the evaluation dimension) is more salient in achromatic stimuli. Since black and white are stronger bipolar opposites than luminous differences in brightness this increases saliency and therefore leads to larger effects. In any case, these results like the results of Study 1 seem to suggest a moderating effect of both stimulus type as well as measurement type on the size of the effect. We will now investigate these notions in more detail as well as to give a more comprehensive estimate of the effect by use of an explorative³ meta-analysis.

We ran a fixed-effects model over the results of both the original study Specker et al. (2018) and our replication. We chose a fixed-effects model rather than a random effects model because the fixed-effects model assumes homogeneity in estimated effect size. Though this assumption often does not hold up when conducting a meta-analysis over a large amount of different studies, the high similarity across our studies made this the best suited model. The results are illustrated in Figure 3. The fixed effects model indicated a meta-analytic effect of 1.31 [1.12, 1.51]. However, there was residual heterogeneity not explained by the model, Q(8) = 86.88, p < .001. This was coherent with previous descriptive comparisons of our included studies which indicated that the small variations of our study design seemed to influence the effect. We therefore ran a fixed-effects model with moderators to see if this was able to explain the residual heterogeneity. As discussed above, our studies employed two different stimulus types (chromatic vs. achromatic) where the achromatic stimuli represent better controlled and more salient stimuli. In addition, when comparing the results descriptively both this replication and the original study reported larger effects when measuring the effect implicitly in comparison to an explicit measurement. Though it has to be mentioned that the smaller effect size is likely due, at least in part, to the fact that reaction time based measures with a lot of trials have less variance and therefore larger effect sizes than a one-trial explicit measure. We therefore used stimulus type (chromatic vs. achromatic) and measure type (implicit vs. explicit) as moderators. The fixed-effects model with moderators showed a significant effect of the moderators, Q(2) = 81.43, p < .001. In addition, there was no residual heterogeneity, Q(6) = 5.44, p > .05, this indicates that all heterogeneity in the effects can be explained by the moderators. The regression coefficient of stimulus type is –1.37, p < .001. This indicates that when stimulus type is chromatic the estimated effect (strength of the association between brightness and positivity) decreases with 1.37. The regression coefficient of measure type is 2.06, p < .001. This indicates that when measure type is implicit the estimated effect increases with 2.06.

We were able to replicate the effects of Specker et al. (2018) in three out of four cases. In addition, the meta-analysis indicated a large effect. In sum, this indicates that there is a strong and replicable association between brightness and positivity. However, the size of the effect is dependent on measurement type as well as stimulus type, as indicated by the meta-analysis. The effect is stronger when measured implicitly than when measured explicitly. As noted, this is likely due to the fact that reaction time based measures with a lot of trials have less variance and therefore larger effect sizes than a one-trial explicit measure. Another likely explanation would be that people confounded “positivity” with “liking”. In other words, if someone saw an image they liked that was dark they may have perceived it as more positive, similarly, when they saw a bright image they disliked they may have perceived it as more negative. In addition, the effect is stronger when using better controlled and more salient stimuli. These findings offer researchers interested in the effect concrete tools when designing a study investigating the effect with regard to effect size estimates for power analysis as well as stimulus and measurement design. Especially researchers who want to use brightness as a cue (e.g., to make the evaluation dimension salient) can use these findings in their research design. To facilitate future research efforts we have made the experimental scripts, analysis scripts, all related stimuli material and our data freely available on the Open Science Framework: https://osf.io/w8479/.⁴

Lakens et al. (2012) have suggested that the association between brightness and positivity may be explained by the strong association people have between black and negativity rather than white and positivity. Due to the relative nature of the IAT one cannot exclude this explanation based on our data. However, the inclusion of chromatic stimuli poses a potential argument against this reasoning since it is based on associations with achromatic stimuli. As noted in Specker et al. (2018), the hue we used (blue), is typically regarded as the most preferred color (e.g., Hemphill, 1996) while being at the same time being stereotypically associated with sadness (Soriano & Valenzuela, 2009). Thus, one can assume blue would be seen as positive (due to being liked) as well as negative (due to its association with sadness). We think it is more likely that people have a general brightness bias as later postulated by Lakens et al. (2013). This more symmetrical relationship is in line with the metaphorical account (Huang et al., 2017; Meier et al., 2004; Meier et al., 2015) where people can represent abstract concepts (such as valence) in terms of concrete concepts (brightness) through metaphoric associations. This notwithstanding, metaphors have to be learned. This is incongruent with the high cross-cultural similarity we found in Specker et al. (2018). In addition, the fact that the effect is larger when using implicit than explicit measures as indicated by the meta-analysis suggest an automatic association that does not rely on explicit metaphors. Admittedly, through learning processes people would be able to internalize explicit metaphorical associations which would lead to implicit associations. Nonetheless, if this were the case one would assume that the original explicit (metaphorical) association would be at least as strong if not stronger than the resulting implicit association. This is not the case, making the metaphorical account less plausible. Thus, our results are in line with accounts such as recently postulated that the association between brightness and positivity is an automatic and universal association that most likely arises early in development.

In sum, the study shows clear support to conclude that there is a strong association between brightness and positivity as well as it providing new insight as to under which conditions the effect is present.

As noted in the main body of the text, our experimental scripts, analysis scripts, all related stimuli material and our data are freely available on the Open Science Framework: https://osf.io/w8479/. The original pre-registration can be found at: https://osf.io/uafyb/.

The writing of this paper was supported by a grant to HL by Wiener Wissenschafts-, Forschungs- und Technologiefonds (WWTF: Projectnumber: CS15-036).

The authors declare to have no competing interests.

Contributed to conception and design: ES

Contributed to acquisition of data: ES

Contributed to analysis and interpretation of data: ES, HL

Drafted and/or revised the article: ES, HL

Approved the submitted version for publication: ES, HL

The author(s) of this paper chose the Open Review option, and the peer review comments are available at: http://doi.org/10.1525/collabra.168.pr