Adjacency beats static and dynamic color in bar chart comparison Authors: Koen Rummens, Sofie Beier Published: 2026 ABSTRACT Bar charts are among the most popular graph formats when communicating quantitative information. Recent work by Boger and Franconeri (2024, Journal of Experimental Psychology: General, 153, 1699–1704) revealed that the spatial design of simple bar charts can strongly alter which relationships viewers detect or miss. In Experiment 1, we replicated their adjacency advantage, showing that misleading relationships in grouped bar charts are more often detected when in adjacent versus non-adjacent (“interleaved”) bars. In Experiment 2, we tested whether color—either static or dynamic—could benefit the detection of interleaved relationships, analogous to the advantage of adjacency in Experiment 1. The two bars containing the misleading relationship were either red or gradually changed from gray to red, with the other two bars in gray. We found no evidence for improved detection with static or dynamic color cues relative to when all bars were uniformly gray. Our findings suggest that the spatial layout of graphs strongly determines which relationships are detected, and reveals limitations of color-based highlighting in guiding our interpretation of bar charts. How to cite this article Rummens, K., & Beier, S. (2026). Adjacency beats static and dynamic color in bar chart comparison. i–Perception, 17(5), 1–10. https://doi.org/10.1177/20416695261481745 Introduction Data visualizations are central to communication and decision-making (Eberhard, 2023) across diverse domains including science (Olowe et al., 2024), business (Dimara et al., 2022), and public policy (Park et al., 2022). Among the many visualization formats, bar charts are especially popular (Borkin et al., 2013), conveying quantitative information in an intuitive and efficient manner. However, as pointed out by Xiong et al. (2024), this efficacy does not extend uniformly to all tasks: Viewers can extract summary values such as mean and variance with little effort, but making comparisons is more effortful and tends to unfold more slowly. Understanding why some data visualization tasks are effortful and others are not requires considering both the viewer's cognitive resources and the visual properties of the graph itself (Shah & Hoeffner, 2002). An important constraint for graph comprehension is the limited capacity of working memory. As readers must keep track of multiple elements such as axes, legends, and data points, complex graphs can easily exceed this capacity. Effective design should therefore minimize working-memory demands (Huang et al., 2009; Lohse, 1997), with a reduced cognitive load often improving judgment and decision-making (see Eberhard, 2023, for an overview). Graph interpretation is also shaped by cognitive biases (Dimara et al., 2018), with viewers relying on mental shortcuts rather than objective features. For instance, the anchoring heuristic can shift data interpretations toward previously encountered and unrelated information (Cho et al., 2017). Together, these findings highlight that graph comprehension is not a neutral process but one constrained by cognitive limitations and biases. In addition to viewer characteristics, the visual properties of data visualizations can also produce unintended effects on graph comprehension. Visual illusions and configural effects induced by graph elements have been shown to distort comparisons between data points (e.g., Amer & Ravindran, 2010; Peebles, 2008). For example, when equally spaced data points are plotted on parallel curves, the vertical distance between them is mathematically identical but often appears unequal to the observer (Zacks & Franconeri, 2020). Even seemingly simple design choices—such as the scaling of the y-axis (Witt, 2019; Yang et al., 2021; Zang & Cousineau, 2025) or ordering of the data (Zhao et al., 2019)—can strongly alter data interpretation. These examples demonstrate that our perception of graphs is not merely driven by the individual data points but also by how these data points perceptually group and interact with the broader context in which they appear. Gestalt grouping principles—such as proximity and similarity—have been proposed to be key determinants of how we perceive and interpret graphs (Ali & Peebles, 2013; Pinker, 1990). Recently, Boger and Franconeri (2024) showed a particularly strong example of how perceptual grouping by spatial proximity can impact observers' interpretation of graph information (see also Talbot et al., 2014). In a bar chart depicting a two-by-two data set, almost half of the participants missed a misleading relationship when the relevant bars were separated by an irrelevant bar (interleaved). Strikingly, all observers detected the anomaly when the relevant bars were adjacent. In the current study, we examined whether color cues can alleviate the cost of spatial separation. Specifically, we asked whether color—either static or dynamic—could improve the detection of relationships in interleaved bars. Our choice to vary color stems from design practice, where color is widely regarded as a primary tool for guiding attention and clarifying relationships in data displays. Large corpora of real-world graphs and charts, such as MASSVIS (http://massvis.mit.edu/), illustrate the widespread use of color in data visualizations. In a 410-graph subset of this repository, more than 66% used at least two or more distinct colors (Borkin et al., 2013). Design guidelines often recommend using color cues to highlight relevant elements, draw attention to anomalies, and group related data (Ajani et al., 2021; Few, 2009; Healey, 1996). Here, given its high perceptual salience (e.g., Drelie Gelasca et al., 2005), we used red to highlight the bars carrying the target relationship. However, despite being commonly applied, static color mapping may only have a modest influence on which comparisons viewers make (Xiong et al., 2024). To provide a potentially stronger attentional signal than static color, we also tested a dynamic cue with the relevant bars changing from gray to red. Such color changes can guide attentional selection (e.g., Lambert et al., 2003; Lu & Zhou, 2005), and are more attention-grabbing than static highlighting in webpage viewing contexts (Das et al., 2024). Yet, how dynamic color cues shape the processing and interpretation of bar charts remains largely unexplored. The present study tests how spatial grouping affects which relationships are more easily detected in grouped bar charts (Experiment 1), and whether color can mitigate—or even abolish—costs of spatial layout in bar chart interpretation (Experiment 2). In Experiment 1, our findings replicated the adjacency advantage reported by Boger and Franconeri (2024), with superior recall for relationships in adjacent versus interleaved bars. In Experiment 2, we investigated whether static or dynamic color highlighting—relative to no highlighting—could enhance the recall of information carried by interleaved bars. Neither static nor dynamic color cues improved performance, suggesting that spatial layout dominates over color in guiding comparisons. Experiment 1: Replication Experiment Method Participants We recruited 200 participants via Prolific (https://www.prolific.com/). All participants were aged between 20 and 45, had proficiency in English, completed secondary education or above, and self-reported having no issues seeing color. Remuneration was 0.70 GBP for the 5-min study. Informed consent was obtained electronically from all participants prior to the start of the experiment. Approval from the Scientific Ethics Committees for the Capital Region of Denmark was not required, as the study does not qualify as a health research project under the Danish Committee Act (Consolidated Act No. 1268 of November 28, 2024, Sections 1(4) and 2(1); https://www.retsinformation.dk/eli/lta/2024/1268). Our protocol complied with the ethical standards of the Declaration of Helsinki and the Danish Code of Conduct for Research Integrity. Stimuli and Materials Stimuli were two variants of a grouped bar chart, both depicting the same 2-by-2 grouped data set (Figure 1A). The data set contains information on the height of two children (River and Charlie) both at age 10 and 12. The misleading relationship in the data is River shrinking between ages 10 and 12. In the first variant of the graph, the misleading relationship is in interleaved or non-adjacent bars (Figure 1A, top bar chart). In the second variant of the graph, the two bars containing the misleading relationship are adjacent (Figure 1A, bottom bar chart). Both the interleaved and the adjacent variant of the bar chart are identical to those used in Boger and Franconeri (2024). Experiments were programmed using Pavlovia Surveys (https://pavlovia.org/; Open Science Tools, Nottingham, UK). Stimuli and results of Experiment 1. (A) The top bar chart in the left panel carries the misleading relationship in interleaved bars. River shrinks between ages 10 and 12, an anomaly which was detected by 72.5% of the observers. The bottom bar chart in the left panel has the misleading relationship in adjacent bars. Detection rate in the interleaved condition improves to nearly 90%. For clarity, target bars are indicated by X's. Participants viewed these bar charts without these markers. (B) Results are summarized in the bar chart in the right panel. The asterisk indicates a significant difference at the .05 level. Error bars represent the bootstrapped 95% confidence intervals. Design and Procedure We varied the spatial proximity of the bars containing the misleading relationship between observers: Half of the participants saw the first variant of the graph (“interleaved condition”), the other half was presented with the second variant of the graph (“adjacent condition”). The procedure was identical to that of Boger and Franconeri (2024). Participants were instructed to remember at least two interesting comparisons or patterns present in the data. As background information on the graph they were about to see, observers read the following text: “Imagine that you are looking at how two kids, Charlie and River, grow over time. Their heights are measured at ages 10 and 12.” Underneath this background, only the axes, but no data, were presented. Before the appearance of the graph, participants could view the axes and labels as long as they wanted. When clicking “Next,” the entire graph—now including the data—was presented for 15 s. The limited presentation duration was communicated beforehand and could be tracked via a countdown timer. After completing two open questions on the patterns they noticed, participants were asked two yes/no questions, the order of which we randomized. The first question asked whether one noticed the misleading relationship actually present in the data (i.e., River shrinking over time). Response accuracy on this question was used to calculate the detection rate of the misleading relationship. The second question was a control question, asking whether a relationship not present in the data (i.e., Charlie's height staying the same over time) was there or not. Participants who incorrectly answered the control question were excluded from the analysis. Analysis and Results After excluding respondents with a false response on the control question, our final sample contained 168 participants. We performed a chi-square test to evaluate the effect of spatial proximity (interleaved vs. adjacent) on the detection rate of the misleading relationship. Confidence intervals (CIs) were estimated separately for each condition using bootstrapping with 20,000 resamples. The detection rate of the misleading relationship in the adjacent condition (89.8%; 95% CI = [0.830, 0.955]) was significantly higher than in the interleaved condition (72.5%; 95% CI = [0.625, 0.825]) (χ2(1, 168) = 8.31, p = .004) (Figure 1B). Specifically, the data were about 9 times more likely under the model where detection rate depends on spatial layout than under the model where both are independent (BF10 = 9.3). Conclusion Our results revealed improved detection of the misleading relationship when in adjacent versus interleaved bars, thereby replicating the adjacency advantage found previously by Boger and Franconeri (2024). Experiment 2: Static and Dynamic Color Cues In Experiment 2, we investigated whether static or dynamic color cues could improve detection performance when the relevant bars were non-adjacent. To this end, we measured the detection rate of the misleading relationship when in interleaved bars. Importantly, the bars containing the misleading relationship were either gray (gray condition), red (red condition) or transitioned from gray to red (gray-to-red condition). To assess any potential benefits of color on performance, the detection rate of the misleading relationship highlighted by a static or dynamic color cue was then compared to performance in the gray condition (where a color cue was absent). Given ceiling performance in the adjacent condition of Experiment 1, we chose not to further investigate the effect of color when target bars were adjacent. Method Participants We recruited 300 participants via Prolific (https://www.prolific.com/). Participants were aged between 20 and 45, had proficiency in English, completed secondary education or above, and self-reported having no issues seeing color. Participants of Experiment 1 were screened out. Remuneration was 0.70 GBP for a 5-min study. Informed consent was obtained electronically from all participants prior to the start of the experiment. Approval from the Scientific Ethics Committees for the Capital Region of Denmark was not required, as the study does not qualify as a health research project under the Danish Committee Act (Consolidated Act No. 1268 of November 28, 2024, Sections 1(4) and 2(1); https://www.retsinformation.dk/eli/lta/2024/1268). Our protocol complied with the ethical standards of the Declaration of Helsinki and the Danish Code of Conduct for Research Integrity. Stimuli and Materials Stimuli were three grouped bar charts depicting the same 2-by-2 data as in Experiment 1. All charts had the misleading relationship in interleaved bars. The first graph was identical to the one in the interleaved condition of Experiment 1 (Figure 2A, top bar chart), with all four bars in gray (#afabaa). In the second variant of the graph, the two interleaved bars carrying the misleading relationship were static red (#ec0a1c), with the two remaining bars in gray (#afabaa) (Figure 2A, middle bar chart). In the third variant of the graph, the two interleaved bars with the misleading relationship gradually changed from gray to red, and the two remaining bars remained gray (Figure 2A, bottom bar chart). Given a viewing period of 15 s, we animated the graph such that the color change began at 2.1 s and completed at 3.9 s. Experiments were programmed using Qualtrics (https://www.qualtrics.com). Stimuli and results of Experiment 2. (A) In the top-left chart of the left panel, the bars carrying the misleading relationship were gray. Detection rate in the gray condition was 73.3%. In the top-right chart of the left panel, the misleading bars were red. The detection rate in the red condition was 69%. In the bottom chart of the left panel, the misleading bars gradually turned from gray to red between 2.1 and 3.9 s. Detection rate in the gray-to-red condition was 75.3%. For clarity, target bars are indicated by X's. Participants viewed these bar charts without these markers. (B) All three interleaved conditions yielded similar detection rates, indicating that neither static nor dynamic color cues facilitated spotting River shrinking from age 10 to 12 (relative to the gray condition). Error bars represent the bootstrapped 95% confidence intervals. Design and Procedure We varied the highlighting of the interleaved bars containing the misleading relationship between observers. Participants were randomly assigned to either the gray condition, the red condition, or the gray-to-red condition. Each condition was completed by 100 unique participants. The procedure was identical to Experiment 1, except for the addition of a manipulation check in the dynamic color condition. After the two yes/no questions, we assessed whether participants accurately perceived the color change. Specifically, participants replied to the following question: “In the beginning of the experiment, I saw for 15 s the graph with (A) all bars in gray—no bars changed color (B) two bars in gray, two bars turning from gray to red once (C) two bars in gray, two bars in red—no bars changed color (D) I don’t recall.” Additionally, to increase clarity for the participants, response options (A)–(C) in the manipulation check were accompanied by the corresponding graph variant (gray, gray-to-red, and red, respectively). The gray-to-red graph was shown as a short repeating animation, so that it remained visually distinct from the static red graph. Analysis and Results A chi-square test revealed no differences between the three interleaved conditions (χ2(2, 255) = 0.86, p = .650). Detection rates were 73.3% for the no-color condition (63/86; 95% CI = [0.640, 0.826]), 69.0% for the static-red condition (58/84; 95% CI = [0.595, 0.786]), and 75.3% for the dynamic gray-to-red condition (64/85; 95% CI = [0.659, 0.835]). A Bayesian contingency-table analysis provided strong evidence for no association between color condition and detection (BF01 = 26.3), indicating that the data were about 26 times more likely under a model where detection does not depend on color condition than under a model where both factors are dependent. Additional Bayesian analyses did not support a detection advantage for static red or gray-to-red highlighting over gray. Specifically, Bayes factors favored no difference over the directional hypothesis that highlighting improves detection (static red: BF0+ = 8.90; gray-to-red: BF0+ = 4.67), suggesting that the observed data are roughly 5 to 9 times more likely under a model with no difference than with a positive effect of color. Responses to the manipulation check revealed that only 61.2% (52/85) of the observers saw the bars changing from gray to red. In total, 23.5% (20/85) of the participants reported that the bars were static red, 11.8% (10/85) reported all bars to be gray, and 3.5% (3/85) did not recall. From those who correctly perceived the bars changing color, 73.1% (38/52) detected the misleading relationship. From those who did not see the color change, 78.7% (26/33) detected the misleading relationship. The detection rates between both response groups did not statistically differ (χ2(1, 85) = 0.354, p = .552). Furthermore, passing versus failing the gray-to-red manipulation check was not associated with higher detection (BF0+ = 6.14); the data were about 6 times more likely under the null model than under a model positing improved performance among those who correctly noticed the color change. Conclusion Neither static nor dynamic color cues improved the detection of a misleading relationship in interleaved bars compared to when color cues were absent. The absence of any meaningful improvement with versus without color-based highlighting underscores the limited capacity of color to mitigate the effect of spatial separation in simple bar charts. Discussion We sought to determine whether color cues could relief costs of spatial separation on the detection of relationships in simple grouped bar charts. In Experiment 1, we showed an adjacency advantage: spatially rearranging the target bars from interleaved to adjacent improved performance. These findings replicated those of Boger and Franconeri (2024), with superior detection of a misleading relationship when the relevant bars are neighboring versus when they are not. In Experiment 2, we highlighted the target relationship in interleaved bars with static and dynamic color cues, but detection did not benefit compared to a no-highlighting condition. Our results are consistent with previous findings revealing the dominance of spatial proximity rather than color in guiding comparisons in grouped bar charts (Bearfield et al., 2024; Xiong et al., 2022). Possibly, other graph formats may elicit different comparison strategies and facilitate the detection of the misleading relationship when in interleaved bars. Specifically, line graphs have been suggested to be well suited to convey information on trends, rather than bar charts, which are thought to be more appropriate for visualizing discrete data (Shah & Hoeffner, 2002). Rather than changes in graph format, refinements to the spatial arrangement of bars may also contribute to effectively guiding attention. For instance, Xiong et al. (2022) showed that vertically aligned bars can promote different comparisons than horizontally grouped ones. Such within-format adjustments should be considered a viable alternative to color cues. In addition to design layout, task characteristics also influence which relationships viewers extract (Matzen et al., 2023). Task demands likely shaped gaze allocation in the current study: While the free recall task of the current study may have favored visually accessible comparisons, a more directed anomaly search might yield different detection patterns. Our findings of Experiment 2 showed that static color mapping did not improve detection of the misleading relationship in interleaved bars: Detection rates were similar when all four bars were gray or when the two relevant bars were static red and the remaining bars were gray. The absence of an improvement in the red condition is consistent with studies emphasizing the importance of spatial proximity in guiding comparisons in bar charts, with minimal effect of static color (Bearfield et al., 2024; Xiong et al., 2022). At the same time, many visualization design guidelines recommend using color grouping or a unique highlight color to direct viewers’ attention to a target relationship and support memory for the intended pattern (Ajani et al., 2021). However, this color-based strategy to increase attentional focus did not appear to work for the simple grouped bar chart in the current study. Despite red being among the most salient colors (Drelie Gelasca et al., 2005), our color manipulation seems to have failed to sufficiently draw attention to the misleading relationship. In the red condition, two perceptual groups were formed by color: a pair of gray bars representing the height of one child (Charlie) and a pair of red bars representing the height of the other child (River). Since each color was equally prevalent, red may not have been distinctive enough to function as a strong attentional cue for the target comparison. One possibility is that color-based highlighting becomes more effective when the target bars are marked by a color that appears less frequently than the color of the surrounding bars. Accordingly, adding additional gray bars (e.g., more children) while keeping only the task-relevant pair in red may increase the distinctiveness of the target relationship and improve its detection despite the greater overall visual complexity. Future work should test this hypothesis by measuring detection rates while varying the relative uniqueness of the highlighted bars. When the relevant bars dynamically changed color in Experiment 2, no improvement in anomaly detection was observed either. This finding suggests that the color change did not provide a stronger cue than static color in guiding attention toward the target relationship. Our manipulation seemed only partially successful, with almost 40% of the observers not seeing the bars change from gray to red. Such change blindness has been suggested to affect graph comprehension, especially when observers are unaware of which aspect of the graph to attend to (Szafir, 2018). Furthermore, cognitive and perceptual load have been suggested to impact change blindness, with better detection under low versus high load (e.g., Lee et al. 2007; Murphy & Murphy, 2018). These factors may have contributed to the change detection misses in the gray-to-red condition, as the basic bar charts in our study still contained a substantial amount of information to be processed in a limited amount of time. Possibly, since gradual color changes can be difficult to notice (Simons et al., 2000), more abrupt or repetitive color changes—rather than the single gradual transition implemented here—may be required for dynamic cues to be accurately perceived. Yet, even when accurately perceiving the color change in the gray-to-red condition, our data suggest no benefit in detection compared to the gray condition. The observers who did perceive the color change (N = 52) observed the misleading relationship in 73% of the cases, which exactly matches the detection rate with all bars in gray. Similarly, when comparing detection rates between the participants who did and those who did not correctly perceive the color change in the gray-to-red condition, our data did not suggest superior detection in the first (73% detection rate) compared to the latter (79% detection rate) group. Overall, our color animation does not seem effective at promoting the comparison of spatially distant bar pairs. In our study, neither the static nor the dynamic use of color succeeded in reducing the cost of perceptual grouping established by spatial proximity. Although we expected our color manipulations to promote different scanning patterns relative to the gray condition, the absence of meaningful detection differences between conditions in Experiment 2 may suggest otherwise. Including eye-tracking in future studies would clarify whether color highlighting was able to direct visual attention or not. More generally, insights in gaze patterns of successful versus unsuccessful detection trials could inform the design of more effective visual cues and support (Schlieder et al., 2025). General Conclusion In the field of design, color has often been considered a primary means of structuring information and directing focus to task-relevant relations. Design guidelines emphasize that appropriate color use can highlight and draw attention to certain data, group and categorize information, and represent magnitudes. From this perspective, color is not only a matter of aesthetics but also a central mechanism for making complex data displays accessible and understandable. Our findings underline a potential gap between design conventions and empirical evidence by demonstrating that, at least in the context of simple grouped bar charts, color cues—whether static or dynamic—are unlikely to improve detection of relationships carried by spatially distant bars. While color is often considered indispensable in design practice, its effectiveness may be limited when competing with strong perceptual tendencies such as spatial proximity. Our work highlights the need for a more nuanced understanding of when and how color can effectively support graph interpretation.