Abstract
In today’s data-saturated society, visual communication has become an essential tool for making complex information accessible. As exposure to visual data increases—whether through educational settings, media outlets, official statistics, public institutions, or social media—the ability to interpret charts becomes a vital form of literacy. However, our knowledge of how charts from real-life contexts are understood and remembered, and what influences their comprehension and recall is relatively modest. This study examines how individuals comprehend and recall information from plain versus embellished bar charts using authentic Eurostat data. In a between-subjects design, participants described chart messages and recalled visuals at two different time points. Unlike traditional comprehension tests, open-ended responses were used to capture spontaneous interpretations, simulating a real-world task: viewing a public statistical chart and making sense of it without guidance. Most participants identified the general topic—often echoing the title—but deeper insights (e.g., pattern recognition, opinions) were rare. While embellishments did not affect message extraction, they significantly improved chart recall. A statistical background positively predicted recall, with some evidence also suggesting an association with comprehension, whereas cognitive reasoning and topic interest showed limited associations. These findings underscore the importance of chart titles, user experience, and design features in shaping understanding. The study advocates for improved chart wording, cautious use of embellishment, and educational efforts to foster data visualization literacy.
Plain Language Summary
Why was this study done?
We are surrounded by charts everywhere. However, little research exists on what people actually take away from them. This study aimed to find out if simple, clean charts are better at communicating information than charts with supporting illustrations. We wanted to understand what helps people grasp charts’ messages and remember them later.
What did the researchers do?
The researchers conducted an online study with 50 people. We showed them eight real bar charts from Eurostat, the European Union’s statistics office. For each topic, one version of the chart was plain, while another was enhanced with illustrations. We asked people to describe the messages of each chart in their own words to see what they naturally focused on. Later, their memory for the charts was tested.
What did the researchers find?
We found a clear difference between understanding and memory:
- Illustrated charts were much easier to remember than plain ones.
- Illustrations did not help people better understand the chart’s messages. People understood the plain and illustrated charts equally well.
- Most people only got a surface-level understanding. While more than 9 out of 10 people could identify a chart’s basic topic, fewer than 1 in 3 spotted deeper trends or patterns.
- People with previous experience in statistics and making/reading charts tended to be better at understanding charts and they were better at remembering the charts.
What do these findings mean?
This study shows that data literacy—the ability to understand and work with data—is a skill that can be learned, not just an innate talent. Our findings suggest that illustrations can be used to make information more memorable, but they are not a shortcut to better understanding. Clear titles and labels remain essential. This research highlights the importance of education and practice to become better at interpreting the visual data.
Introduction
In today’s data-saturated society, visual communication has become an essential tool for making complex information accessible. As visual content better attracts people’s attention compared to words (Paivio & Csapo, 1973; Pettersson, 1994), it can be particularly useful for communicating data. Data visualizations transform abstract numbers into concrete visuals that our perceptual system can process quickly and efficiently (Ware, 2004). Data visualizations have, therefore, been widely used to inform the public and engage audiences. As exposure to visual data increases—whether through educational settings, media outlets, official statistics, public institutions, or social media—the ability to interpret charts becomes a vital form of literacy. However, knowledge about how charts from real-life contexts are understood and remembered, and what influences comprehension and recall is relatively modest.
First, research is scarce on the cognitive processes involved in the interpretation of charts. A topic that is well studied is how people extract numerical values from charts; several researchers have investigated graphical elements to identify those that best communicate numbers (e.g., length better than area or volume; Cleveland & McGill, 1984; Spence, 1990). On the contrary, how individuals construct meaning from charts, the (overall) message, or even a story, to be communicated has been emphasized only lately. While recent popular and design-oriented literature has highlighted the significance of effective communication (e.g., Camões, 2016; Knaflic, 2015; Schwabish, 2021), empirical studies of how people spontaneously interpret such messages are still sparse. This is especially true for charts depicting unfamiliar or official data that lack explicit explanatory text. Our study addresses this gap by examining how users interpret charts, using the concept of chart messaging—defined here as insights or meanings that users take away from the chart, intended to go beyond simply reading off numbers. In addition, we introduce the concept of title clutching, referring to the tendency of users to repeat or paraphrase the chart title instead of generating an independent chart interpretation.
Furthermore, visual retention and the cognitive processing of charts have been linked to a prominent dilemma—with rather mixed evidence—whether charts should have a minimalistic, de-cluttered design, as opposed to embellished design. In this study, embellishment refers to the addition of pictorial elements (e.g., illustrations, icons, images, textures) that support the meaning, information, or narrative visually, and thus go beyond the graphical components needed to display the data. These design choices raise important questions about the trade-off between aesthetic appeal and cognitive clarity. While some studies suggest embellishments help with memory and engagement (Bateman et al., 2010), others argue that they distract from the data or reduce accuracy (Few, 2012). This ongoing debate often lacks ecological validity because studies are conducted in artificial settings or use stylized stimuli. Given the volume of information competing for public attention today, this dilemma bears even broader implications. Our study contributes to this discussion by examining user responses to authentic public data charts, thereby simulating real-world encounters with statistical information.
Additionally, individual differences—such as education, statistical experience, cognitive ability, and topic interest—are likely to shape chart comprehension, but are frequently overlooked in empirical research on data visualization. By incorporating measures of reasoning ability, statistical background, and personal interest, our study also aims to explore how user characteristics interact with design features to shape comprehension and recall.
Second, methodologically speaking, prior research has often evaluated chart comprehension through accuracy and speed of data processing or preference ratings. Most existing research has been conducted in artificial settings where participants responded to predefined questions. Although such controlled conditions are beneficial for investigating discrete cognitive processes and ensuring experimental precision, they inherently create a detachment from naturalistic situations, thereby limiting the external validity and real-world applicability of the findings. Therefore, studies are lacking that examine how visual design choices affect the process of interpretation in authentic settings where users must make sense of unfamiliar data without guidance.
To summarize, data visualizations are ubiquitous in everyday contexts but are rarely accompanied by guidance on how to read them. Relatively little is known about what people really take away from data visualizations in real life. Existing studies have often focused on perceptual accuracy or design features, yet less is understood about how audiences interpret visualizations, what they retain, and how these interpretations influence their understanding of issues or decisions. To fully understand the effectiveness of visual communication, it is thus not sufficient to test whether people can extract specific numbers quickly and accurately; it is equally important to examine what messages they grasp from the charts. The current study addresses this gap by exploring spontaneous comprehension and recall while simulating a real-world task—viewing public statistical charts and trying to make sense of them without guidance—and controlling for user characteristics. This rationale positions our study at the intersection of cognition and communication. The study is structured around three research questions:
to explore how people understand what charts want to communicate to them: What messages do the users extract from charts? (RQ1)
to investigate the role of embellished versus minimalistic chart design: How does the visual design (plain vs. embellished) relate to chart messaging and recall? (RQ2), and
to account for other potential factors that contribute to individual differences in shaping chart comprehension and recall: How do background characteristics (education and experience, reasoning skills and interest) relate to chart messaging and recall? (RQ3).
Methodologically, we aim to contribute to the growing body of evidence on chart interpretation by using a mixed-methods approach, combining experimental and qualitative methods. Using authentic statistical charts published by Eurostat, we investigate how users engage with data visualizations in contexts that mirror everyday information consumption.
The paper first reviews the literature on relevant aspects of data visualization research and then presents the methods used in the study. The results section is divided into two parts: the qualitative part on messages extracted from the charts (RQ1) and the quantitative part on factors (visual design—RQ2 and individual characteristics—RQ3) associated with messaging and recall. The paper concludes with discussion and conclusion.
Theoretical and Empirical Background
This chapter provides the theoretical and empirical background for the study. It presents the theoretical framework, guidelines for charts in visual communication, and research specifically relevant for the three research questions: research on how people extract meaning from charts (RQ1), research on the role of visual design in chart comprehension and recall (RQ2), and research on individual characteristics that may shape users’ interpretation of data visualizations (RQ3).
Theoretical Framework
Data visualization research has a strong interdisciplinary foundation. Its theoretical background draws on findings from statistics, cognitive science and psychology, computer and information science, semiotics, communication, graphic design, and educational research. To understand how individuals comprehend and recall chart messages, this study builds on Schema Theory (Pinker, 1990) and Cognitive Load Theory (Lohse, 1993; Paas et al., 2003). Schema Theory proposes that comprehension relies on acquired mental structures, or schemas, that organize prior knowledge and guide interpretation (Pinker, 1990). In the context of data visualization, chart schemas allow users to map graphical conventions (e.g., bar height corresponding to magnitude) onto meaning, while Kosslyn (2006) emphasizes the interaction between perceptual constraints and higher-order cognitive processes.
This work is also informed by Dual Coding Theory (Paivio, 1971, 1975), which suggests visual representations support comprehension when paired with verbal or symbolic encoding. Cognitive Load Theory further suggests that human working memory has limited processing capacity (Paas et al., 2003). Charts require viewers to simultaneously process visual encodings, compare values, and integrate semantic context (Lohse, 1993). Excessive embellishments or complex designs can increase extraneous load, distracting from the intended message. Conversely, minimalistic charts may reduce cognitive load, but risk omitting features that aid memorability and engagement. Thus, design choices directly affect the balance between cognitive effort and communicative clarity.
Charts in Visual Communication
Building on this theoretical background and on the experiences of practitioners (e.g., Camões, 2016; Knaflic, 2015; Schwabish, 2021), several practical guidelines for designing effective charts have been proposed. These include favoring simple, conventional encodings that enable viewers to quickly grasp intended messages; choosing familiar chart types to leverage existing knowledge structures; and providing titles, labels, and scales that clearly communicate what the chart represents while guiding interpretation. Equally important are practices to avoid in order to prevent or reduce bias, confusion, and misinterpretation. These include chart clutter (e.g., excessive gridlines, legends, 3D effects, or decorative elements); broken axes or inconsistent scaling; and excessive embellishment, which can overwhelm the message and increase cognitive load.
Although selecting an appropriate chart type and following visualization guidelines can support comprehension (Hegarty, 2011; Hehman & Xie, 2021; Kosslyn, 2006; Shah & Hoeffner, 2002), chart interpretation is not always straightforward. Users may assume that charts are objective or neutral, especially when they appear in textbooks, news articles, or official reports. Yet, visualizations are not immune to bias (Lin & Thornton, 2021; Tal & Wansink, 2016). Ineffective or counter-intuitive designs that disregard established conventions can exacerbate bias, leading to deception, errors, and misinterpretation (Correll et al., 2020; Franconeri et al., 2021; Lauer & O’Brien, 2020; A. V. Pandey et al., 2015; Szafir, 2018; Wijnker et al., 2022; Yang et al., 2021). In other words, design decisions that deviate from conventions increase the risk of confusion or deception. This is particularly critical when communicating trending or controversial topics, such as global climate change, politics, public policy, or public health (Cairo, 2019; Lee et al., 2021).
While many different chart types exist, our study focuses on bar charts because they are widely applicable, familiar, and frequently used. Although primarily designed for categorical variables, bars can also represent numerical variables (e.g., time series). In a bar chart, an acquired schema signals that larger values correspond to proportionately higher bar tops (Franconeri et al., 2021; Pinker, 1990). This schema is probably highly familiar to most users. However, when encountering a novel or unconventional chart type (e.g., a radar chart, a chord chart, a triangular graph), individuals must acquire new schemas to support interpretation.
Chart Comprehension: Cognitive and Communicative Processes (RQ1)
Data visualization research has shifted its focus from examining the impact of visual channels (or pre-attentive attributes) such as position, size, length, area, angle, shape, and color intensity (Cleveland & McGill, 1984; Spence, 1990) to additional factors that shape chart comprehension. The current view transcends visual perception of graphical elements and recognizes a complex sequence of underlying cognitive processes, where several mechanisms can be observed. These include identification, comparison, and segmentation of objects and trends, gist understanding, and reasoning about semantic content (Bertin, 1981; Burns et al., 2020; Franconeri, 2021; Franconeri et al., 2021; Gleicher et al., 2011; Hegarty, 2011; Shah & Hoeffner, 2002; Szafir et al., 2016; Tufte, 2001; Ware, 2004). Extracting information from charts typically involves comparisons, suggesting that the limited capacity of working memory may hinder comprehension. However, our understanding of the specific cognitive mechanisms involved in chart comprehension remains incomplete.
Given the communicative purpose of charts, understanding how people make sense of these visual representations is crucial. While early research relied heavily on perceptual and psychophysical testing, more recent studies have adopted qualitative and mixed-methods approaches to explore how users interpret visualizations (North, 2006). Such approaches provide more nuanced insights into how viewers construct meaning from charts (Lauer & O’Brien, 2020; Lundgard & Satyanarayan, 2021; C. Stokes et al., 2023; L. Stokes & Hearst, 2022). These findings underscore that chart comprehension extends beyond perceptual decoding and involves diverse interpretive, inferential, and contextual processes. An influential contribution is the framework proposed by Lundgard and Satyanarayan (2021), which identifies four levels of semantic content. This framework offers a valuable lens for analyzing how users interpret charts—from basic recognition of graphical elements to complex domain-specific insights:
Level 1: Enumerating visualization construction properties, such as marks and encodings (e.g., describing the type of chart, axis labels, and scales).
Level 2: Reporting abstract statistical concepts and relationships within the dataset (e.g., descriptive statistics like mean, standard deviation, and point-wise comparisons).
Level 3: Identifying perceptual and cognitive phenomena, such as conveying the overall gist of complex trends and patterns.
Level 4: Elucidating contextual and domain-specific insights, such as social and political context, which depend on the reader’s background knowledge and expertise, as well as external factors influencing the depicted data.
This framework is particularly valuable for analyzing open-ended user responses, as it captures both literal chart reading and higher-order interpretive reasoning. It reflects the broader shift in visualization research from a focus on visual channels to an emphasis on cognitive and communicative mechanisms, thereby bridging the gap between perceptual accuracy and real-world comprehension.
Minimalism Versus Embellishment (RQ2)
A longstanding dilemma in chart design concerns the choice between a minimalistic, de-cluttered style and an embellished display, sometimes referred to as chartjunk in extreme cases. Extant guidelines mainly recommend maximizing the “data-ink ratio” (Tufte, 2001) and avoiding chart “clutter” in the form of excessive gridlines, legends, 3D effects, color overload, and other pictorial elements (e.g., Camões, 2016; Cleveland, 1985; Few, 2012; Hehman & Xie, 2021; Knaflic, 2015; Kosslyn, 2006; Schwabish, 2021). Some empirical evidence supports these recommendations (Ajani et al., 2021), while other findings are inconclusive and context dependent (Gillan & Richman, 1994; Levy et al., 1996; Skau et al., 2015), or even conflicting. Users often prefer embellished charts, likely because they find them more familiar (Hill et al., 2017; Inbar et al., 2007) or due to habituation to such displays (Fisher et al., 1997). Embellished charts are also better memorized (Borkin et al., 2013) and recalled both in the short term (Borkin et al., 2016; Li & Moacdieh, 2014) and long term (Bateman et al., 2010), and they may also increase engagement (Haroz et al., 2015). This debate can be interpreted through the lens of cognitive theory. Minimalist designs may reduce extraneous cognitive load and allow for more precise data extraction. However, embellishments may foster dual coding by adding meaningful visual associations, thereby supporting recall.
Individual Differences and Chart Comprehension (RQ3)
Despite theoretical insights, empirical studies have often overlooked how individual differences influence chart comprehension. Prior research showed that education plays a crucial role in developing data visualization and communication skills, which are important for efficient citizenship in today’s society. For example, students frequently encounter difficulties comprehending textbook charts (Shah & Hoeffner, 2002). We understand charts once we learn them and become familiar with their format. Chart comprehension requires a graph schema, acquired knowledge, and expectations about how charts function (Franconeri et al., 2021; Pinker, 1990). Without such schemas, users may struggle to interpret visualizations or may instead rely on superficial cues such as titles. This mechanism helps explain why educational background and prior experience can affect chart comprehension.
Knowledge of chart conventions is typically taught in statistics courses but can also be developed through practical experience with data visualization. Such knowledge forms part of statistical literacy, defined as the ability to interpret, critically evaluate, and communicate statistical information (Gal, 2002; Gal & Ograjenšek, 2017). Beyond knowledge, cognitive abilities may impact chart comprehension. These include pattern recognition (Lohse, 1993), general intelligence, and abstract reasoning (Duncan et al., 2017). Moreover, research in educational psychology suggests that individuals are more likely to invest cognitive effort in topics they find personally interesting (Krapp, 1999). Finally, chart comprehension also requires knowledge of the underlying content (Hegarty, 2011; Shah & Hoeffner, 2002).
Based on these considerations, the present study was designed to empirically examine knowledge gaps in chart comprehension and recall using real-world statistical visualizations and a mixed-methods approach, with the aim of better understanding how data design features (plain vs. embellished) and individual differences (background, reasoning ability, and interest) interact to shape the comprehension and recall of public data communication.
Method
To investigate how users interpret and recall information from plain and embellished bar charts, we used a mixed-method approach. Rather than relying on multiple-choice items to assess comprehension, we followed calls to “give voice” to users (North, 2006) and analyzed free-text chart descriptions. The measurement instrument asked each participant to describe eight charts in their own words (qualitative data for RQ1) and collected data on participants’ characteristics (quantitative data for RQ3). Participants were randomly assigned to two groups, each presented with four charts in the plain format and four embellished charts in random order (an experiment for RQ2).
Participants
The sample consisted of 50 participants, gender-balanced, with a mean age of 32.02 years (SD = 11.38, Min = 18, Max = 71). The study was conducted on the Prolific research platform in December 2021, in line with the trend of using crowdsourcing platforms in data visualization research (Borkin et al., 2013, 2016; Heer et al., 2010). Eligibility criteria were applied to ensure survey response and exposure to data visualizations (reliable past response behavior, good vision, prior management experience).
Participants were randomly assigned to two groups, with the goal of achieving equal group sizes. However, due to participant dropout, the final group sizes were slightly unbalanced (n1 = 27, n2 = 23).
The study did not collect any sensitive data and the participation was voluntary. The respondents had the right to withdraw at any point during the study. We estimated the time needed for response to approximately 40 min. Participants were paid 5.40 GBP per completed questionnaire which was above the average payment rate on Prolific at the time.
Material and Procedure
The measurement instrument builds on an earlier testing study, was administered using Qualtrics and included a series of visual and cognitive tasks. It was pre-tested in-depth with four people and their thorough feedback contributed to the final version of the instrument, including the elimination of two charts due to the pilot instrument taking longer than expected.
We used bar charts comparing the European Union Member States that Eurostat posted on Twitter. For each of the eight topics (household consumption expenditure on alcoholic beverage; use of cloud computing in enterprises; severe material deprivation rate; protected terrestrial area; road traffic victims; influenza vaccination rate; municipal waste generated; and women in national parliaments), two formats were available: a plain bar chart and an embellished version with the same indicator (so, 16 charts altogether; all charts are provided in the measurement instrument, see Data Availability Statement). While the embellished and plain charts represented the same indicators, the data came from different years. Two groups of participants were each presented with eight charts of different topics in random order—four were in the plain format and four were embellished. Figures 1 and 2 illustrate the two formats for the topic of women in national parliaments. The plain chart (Figure 1) is a conventional vertical bar chart with a white background. It displays the proportion of seats held by women in national parliaments along the vertical axis, which extends to 50% to indicate gender parity. The bars represent countries in decreasing order, with EU member states shown separately from other countries of the European Economic Area. All countries are depicted with green columns, except for the EU average, which is blue. The embellished chart (Figure 2) contains two key embellishments: the figure of a woman at a podium and a semi-circular form of the bar chart with a red background, designed in the style of a parliamentary chamber. Here, the vertical axis extends to 100% but emphasizes 50% to indicate gender parity. The UK is presented separately and the label for the EU average is in a different color.
Example of a plain bar chart showing the proportion of seats held by women in national parliaments across European countries in 2018, with the EU treated as a group and its average highlighted.
Example of an embellished semi-circular bar chart showing the proportion of seats held by women in national parliaments across European countries in 2019, designed in the style of a parliamentary chamber and featuring the figure of a woman at a podium, with the EU treated as a group and its average highlighted.
Participants first rated their interest in each of the eight chart topics. Then, each participant viewed eight charts (four plain, four embellished) in random order and responded to an open-ended question requesting description of chart messages in as much detail as possible. Following this, participants completed two memory recall tasks: Memory test I (first recall of charts immediately after viewing and describing all charts) and Memory test II (second recall of charts after a set of intervening tasks). In between the first and second memory test, participants responded to socio-demographic survey questions regarding their education and experience with charts and completed a timed cognitive reasoning test, the short Hagen Matrices Test (HMT-S).
Measures
Dependent Variables
Messaging was derived from coded qualitative charts descriptions. For each message or code identified in a participant’s response (n = 400), they were awarded one point. Messaging scores for each chart thus ranged from 0 to 6 (Min = 0, Max = 5, M = 2.57, SD = 1.20), reflecting the extent of information participants extracted from the chart. As an alternative operationalization, we also counted the number of words in each chart description (Min = 5, Max = 114, M = 31.21, SD = 21.25) to assess whether this simpler variable could serve as a proxy for the more labor-intensive qualitative analysis.
Recall variables derived from Memory tests I and II. At the level of a respondent (n = 50), we calculated the number of charts mentioned in each test. The two variables, thus, ranged from 0 to 8 (Min = 0, Max = 8, M = 4.22, SD = 2.26 for the first recall; Min = 0, Max = 8, M = 4.68, SD = 1.96 for the second recall). At the level of a chart (n = 400), we dummy-coded variables indicating whether a chart was recalled (1) or not (0) in each test.
Independent Variables
Chart format
Each chart was labeled either plain (coded 0) or embellished (1)
Interest in topic
Self-reported interest in each chart topic was measured on a 5-point scale (ranging from “1 = not interesting at all” to “5 = extremely interesting”). Average interest ranged between 2.62 (for the chart on material deprivation rate, n = 50) and 3.38 (for the chart on women in national parliaments, n = 50).
Title clutching index
We controlled for clutching on the words in the chart title by calculating the share of the words from the title in respondents’ descriptions (with omission of the conjunctives): M = 0.17 (SD = 0.12, Min = 0.00, Max = 0.69, n = 400). As the chart title typically contributes to just one message (coded about in our analysis, for details see Results section), we expected that a respondent would get fewer messages if relying more on the title.
Background
A composite background score (0–4) was based on four, dichotomously coded, variables: bachelor’s degree or higher, learned about statistics, some/lots of experience reading charts, and some/lots of experience creating charts: M = 2.56 (SD = 1.26, Min = 0, Max = 4, n = 50). 70% of participants reported having at least bachelor’s degree, 82% learned statistics, 56% were with some/lots of experience in reading charts, while 48% with some/lots of experience in creating charts.
Cognitive reasoning score
The score was obtained from the short version of the Hagen Matrices test HMT-S (Heydasch, 2013). The test comprises six problems where the participant has to identify the pattern in incomplete 3 × 3 matrices to select the correct missing part among eight options. The test is time-limited, with participants having 2 min to solve each problem, a fact that was communicated to them in advance. A visible timer was provided during the tasks.
The score ranged from 0 to 6 with M = 3.78 (SD = 1.39, Min = 1, Max = 6, n = 50). Although the test has shown acceptable reliability in prior studies (KR20 = 0.62; Heydasch, 2013), reliability in our sample was lower (KR20 = 0.46), limiting the ability to draw definitive conclusions about abstract thinking abilities.
Analysis
Data were analyzed at two levels: the participant level (n = 50) and the chart level (n = 400 chart descriptions). All qualitative responses were analyzed in Excel using an inductive coding approach, with no pre-existing categorization scheme applied. The process began with open coding, in which responses were carefully reviewed and segments of text that conveyed a specific message were assigned descriptive labels. Through an iterative process of comparison and refinement, the initial fragmented codes were consolidated into a final set of six message codes. Both authors contributed to this process by coding the responses independently, collaborating to develop the coding scheme, and resolving discrepancies in borderline cases through discussion until consensus was reached. The coding was data-driven, aiming to remain as close as possible to participants’ own words and meanings, and ensuring that the resulting categories were grounded in the empirical material rather than imposed a priori. Finally, the codes were summarized using a content analysis framework, which enabled systematic categorization and quantification of message types (Coe & Scacco, 2017). Subsequently, the coding scheme was contrasted with Lundgard and Satyanarayan’s (2021) framework to assess conceptual alignment.
To evaluate how independent variables predicted two outcome variables—messaging and recall—we conducted linear mixed-effects modeling for messaging scores and binary logistic regression for recall. Independent variables were first tested individually, then included in multivariate models, adhering to the rule of 10 observations per predictor variable. We applied linear mixed models to account for multiple observations on the same participants (McCulloch & Searle, 2000; Seltman, 2018). To estimate the association between independent variables and messaging, we used a model with random intercept to account for within-subject correlation. For the dummy-coded outcome of memory recall, we used binary multiple logistic regression.
All quantitative analyses were conducted using IBM SPSS Statistics 25 and R software (nlme and jtools packages). No missing data were observed, and no imputation was required.
Results
Messages Extracted From the Charts
Each participant described eight charts in as much detail as possible in order to understand what messages they extracted from charts (RQ1). Coding of qualitative participants’ descriptions identified six messages extracted from the charts (n = 400; see Table 1). Almost all chart descriptions included a basic description of the chart’s subject matter. About two fifths of the descriptions observed extreme values or measures of central tendency. Observations of patterns and affective expressions surfaced in less than a third of descriptions; least common were technical comments. Mapping these findings onto Lundgard and Satyanarayan’s (2021) framework, we observe parallels: the Other category aligns with Level 1 (enumerating visualization construction properties), while Extremes and Average reflect Level 2 (reporting abstract statistical concepts and relationships). Patterns correspond to Level 3 (identifying perceptual and cognitive phenomena), and Affective expressions may relate to Level 4 (elucidating contextual and domain-specific insights).
Messages Extracted From Charts (n = 400).
Note. The same sentence could be coded with more than one code.
Furthermore, the analysis identified two common errors in chart interpretation, namely misunderstanding the presented indicator (e.g., “the percentage use of cloud computing per enterprise” instead of the percentage of enterprises that use cloud computing) and mistaking relative for absolute values (e.g., “It tries to show which country has more terrestrial natura than others.” vs. which country has a larger share of terrestrial natura). Here it should be noted that follow-up questions are required to determine whether the problem lied in the lack of understanding or difficulty in expressing the understanding.
Factors Affecting the Messaging
In this section, we focus on the parts of research questions RQ2 and RQ3 that relate to messaging. The linear mixed model with the messaging score as the dependent variable is presented in Table 2. Assessed VIFs were as follows: 1.05 for chart format, 1.01 for interest, 1.05 for title clutching index, 1.10 for background, and 1.10 for matrices score. Hence, no multicollinearity issue was detected. Although correlation between the messaging score and title clutching index was negligible (Kendall’s tau = −0.15, p < .001, n = 400), the model showed the title clutching index to be a significant predictor (Est = −0.97, p = .02). Greater title clutching predicted lower chart messaging. We also found a positive main effect of educational and experience background, albeit very small and not statistically significant (Est = 0.21, p = .07). We observed only moderate correlation between background and matrices performance (Kendall’s tau = 0.21, p = .07, n = 50) and the latter did not significantly predict the messaging score (Est = 0.04, p = .71). The chart format and the interest in the topic were not significantly associated with the messaging score.
Linear Mixed Model for the Messaging Score.
The messaging score was correlated with the number of words used in chart description (Kendall’s tau = 0.61, p < .001, n = 400). If we replaced the messaging score with the number of words as the dependent variable, we found two significant predictors for the linear mixed model, namely the negative effect of title clutching index (Est = −17.81, 95% CI [−28.37, −7.25]; t(347) = −3.32, p = .001) and the positive effect of background (Est = 4.34, 95% CI [0.04, 8.63]; t(47) = 2.03, p = .05).
Factors Affecting Recall
In this section, we focus on the parts of research questions RQ2 and RQ3 that relate to chart recall. We first ran several paired samples t-tests (with alpha cut-off of p = .017 adjusted for Bonferroni correction for multiple testing). The results showed that the number of recalled charts was lower at the first recall (M = 4.22, SD = 2.26) compared to the second one (M = 4.68, SD = 1.96), t(49) = −2.39, p = .02 (95% Mean Difference CI [−0.85, −0.07], Cohen’s d = 0.20). The results also showed that at the first recall, participants recalled significantly fewer plain charts (M = 1.70, SD = 1.28) than embellished ones (M = 2.60, SD = 1.25), t(49) = −5.23, p < .001 (95% Mean Difference CI [−1.25, −0.55], Cohen’s d = 0.70). A significant effect of the chart format was also observed at the second recall (M = 1.94, SD = 1.22 and M = 2.74, SD = 1.01 for plain and embellished charts, respectively, t(49) = −5.29, p < .001 (95% Mean Difference CI [−1.10, −0.50]), Cohen’s d = 0.72).
The binary multiple logistic regression model is summarized in Table 3. We composed a dummy-coded variable, indicating whether the chart was recalled or not. We did not detect multicollinearity issues. Assessed VIFs for Memory test I were as follows: 1.03 for chart format, 1.02 for interest, 1.06 for title clutching index, 1.16 for background, and 1.14 for matrices score, while for Memory test II we found the following VIFs: 1.02 for chart format, 1.02 for interest, 1.07 for title clutching index, 1.16 for background, and 1.16 for matrices score. The results show that, holding all other predictor variables constant, the odds of a chart being recalled was significantly higher if the presented chart was embellished, both for the first recall (95% CI [1.58, 3.64]; p < .001) and second recall (95% CI [1.38, 3.22]; p < .001). We also found that recall odds increased with education and experience background (95% CI [1.11, 1.58]; p = .002, and [1.02, 1.46]; p = .03, for the first recall and second recall, respectively). For the second recall, odds decreased with clutching on the title (95% CI [0.01, 0.40]; p = .004). Results suggest that higher scores on the matrices test decreased the odds for the first recall (95% CI [0.69, 0.94]; p = .01) as well as for the second recall (95% CI [0.61, 0.85]; p < .001).
Binary Multiple Logistic Regression Model for Chart Recall.
Discussion
Bar Charts: The Challenge of Deeper Comprehension (RQ1)
Our first research question examined the level of engagement with authentic bar charts. As expected, most participants achieved a surface-level understanding, typically echoing the chart’s title or providing a basic content description. This confirms the critical role of the chart title in shaping initial interpretation. The messages that could indicate more engagement and further cognitive processing, so deeper comprehension—for example, searching for patterns, drawing inferences, or expressing opinion—were far less common. This large gap between recognizing the chart topic and extracting richer meaning suggests that users may be getting much less out of even common charts than hoped for, particularly given that our sample was relatively highly educated. A few descriptions hinted at potential misunderstandings of visualized indicators, which further highlights the difficulty in interpreting chart content and the need to move beyond simple accuracy metrics when assessing comprehension.
Embellishment, Recall, and the Chartjunk Debate (RQ2)
Our results contribute a nuanced perspective to the ongoing debate regarding minimalistic design, embellishment, and “chartjunk” (Bateman et al., 2010; Borkin et al., 2013, 2016; Tufte, 2001). We found that embellishing the bar charts did not affect message extraction (messaging) but significantly aided recall. Participants recalled embellished charts better than plain ones. This result aligns with previous studies that report a positive effect of embellishment on memory (Bateman et al., 2010; Borkin et al., 2016), while supporting research that shows no effect on accuracy (Bateman et al., 2010).
By separating messaging (comprehension) and recall (memory), our results suggest that embellishments may serve different cognitive functions: they support memorability by directing attention without necessarily enhancing the depth of understanding. This offers a partial reconciliation of conflicting prior evidence. Minimalism may help prevent distraction when the goal is extracting detailed information, but our results provide adequate evidence that strategic embellishments can be considered for public-facing charts where memorability is a key goal.
Contrary to the expected impact of temporal decay of short-term memory (Cowan, 2008), we found an increase in the number of recalled charts between the first and second recall tests, which we attribute to the prompting and learning-enhancing effect of the repeated survey question that likely offset the expected temporal decay of short-term memory.
The Power of Background and Acquired Skills (RQ3)
Our third research question investigated the role of participant characteristics. Our results show that both chart messaging and recall tend to be positively related to background, a variable describing a cumulative score of education and experience with statistics and charts, even if the background did not meet the conventional standard for significance when predicting messaging. This finding aligns with the notion that data visualization literacy is an acquired trait that improves with relevant experience. This directly contrasts with some prior inconclusive (A. V. Pandey et al., 2015), no effect (Lauer & O’Brien, 2020), or negative findings (Lin & Thornton, 2021), possibly due to our use of a more experienced, non-student sample.
However, the role of other factors was less clear. The matrices score of cognitive reasoning ability did not predict chart messaging, suggesting that general familiarity with bar charts might mitigate the advantage of higher abstract reasoning skills. Additionally, a negative association between higher matrices score and the odds for chart recall could reflect different cognitive abilities: the matrix test evaluates abstract pattern recognition, while chart recall relies on verbal articulation. The lack of a clear relationship for cognitive reasoning, however, warrants caution, as the reliability of the matrix test in our sample was not confirmed, limiting the interpretation of these particular findings. Furthermore, we found no significant relationship between pre-assessed interest in the topic and either chart messaging or recall. This indicates that while background experience may be a predictor of performance, basic engagement with a visualization may occur even when initial topical interest is low, reinforcing the potential power of data visualization to captivate and inform audiences.
Implications for Design and Education
The totality of our findings suggests two sets of recommendations for practice. The first set of recommendations refers to chart characteristics and design. Our results suggest a nuanced design approach: while embellishments can be strategically employed to boost memory, they are insufficient to ensure deeper understanding.
The central role of the chart title in messaging is reaffirmed. However, “clutching” to the exact title wording appeared associated with less thorough messaging and lower odds for chart recall. Conversely, users who employed their own vocabulary to paraphrase the title provided richer chart descriptions and recalled more charts. This aligns with past research showing people recall charts by rewording the title (Borkin et al., 2016). Interestingly, few participants explicitly acknowledged relying on titles when interpreting charts (Lauer & O’Brien, 2020). Nevertheless, careful attention should be paid to title wording, as it can influence comprehension—particularly when communicating controversial topics (Kong et al., 2019). Moreover, the errors observed in interpretation of visualized indicators call for improved title phrasing and testing to minimize miscommunication.
Further research should clarify whether the importance of title words stems merely from their prominence and whether other textual elements built in charts could attract more attention if equally prominent. Various annotations could help with both correct interpretation and passing on more messages, though designers must exercise caution to avoid adding unnecessary chart clutter. Recent studies emphasize the challenges and opportunities in combining textual and visual elements to improve interpretation and underscore the importance of thoughtful integration, demonstrating that readers often benefit from additional explanatory text (Stokes and Hearst, 2022; C. Stokes et al., 2023). These findings suggest that chart comprehension is enhanced when visual displays are supported by well-integrated written elements. Finally, the increasing role of AI offers a promising avenue for advancing chart comprehension, as AI tools could potentially help generate optimized chart titles and annotations, facilitating more effective and engaging visual communication.
The second set of recommendations refers to education and training. Data visualizations are central tools for communicating scientific and statistical information. To support the development of chart comprehension, educators must understand how learners interpret visual information and how teaching practices can scaffold the acquisition of these skills. Our findings suggest that both messaging and recall in common bar charts might be more dependent on “acquired” traits (education and experience) rather than “given” ones (cognitive reasoning). This finding suggests a thorough evaluation of the extent to which current educational curricula address data visualization, adding another piece to the already established line of calls for dedicated data visualization education (Lauer & O’Brien, 2020; Lin & Thornton, 2021).
Limitations of the Study
This study offers an exploratory investigation into how individuals comprehend and recall information from different types of bar charts. While our findings highlight some promising trends, several methodological limitations should be acknowledged when interpreting the present findings. First, the relatively small sample size (n = 50) reduces statistical power and limits the generalizability of the results, making it difficult to isolate the effects of visualization type from individual differences. While meaningful patterns were observed, future studies with larger and more diverse samples are needed to strengthen confidence in the findings and to test their robustness across different populations.
Second, our measure of cognitive ability (HMT-S) showed poor internal consistency (KR-20 = 0.46). This low reliability likely introduced measurement error and increased variability, which may have attenuated the observed associations of cognitive ability with messaging and recall. Our choice of the HMT-S was driven by our focus on reasoning and pattern recognition, constraints of online administration and task duration, but future studies should incorporate more reliable cognitive assessments (e.g., the full-length HMT, which may offer greater internal consistency, or the well-established Raven’s Progressive Matrices) and provide clearer rationales for test selection.
Third, although efforts were made to ensure comparability between chart stimuli, the plain and embellished versions were based on data from different years as we wanted to only use authentic charts. This difference introduces a potential confound, as observed effects could partly reflect year-to-year variation in the data rather than purely design-related influences. In addition, stimuli characteristics (e.g., title lengths and visual salience of key elements) may have inadvertently influenced results beyond the intended manipulation of visual design. Replicating the study with identical datasets across conditions would help address this concern.
Fourth, participant background information, such as prior experience with charts was measured through self-report. Self-report measures are susceptible to response bias and measurement error, which may limit their ability to adequately reflect actual knowledge and skills. Future studies could benefit from using validated instruments, such as Graph Literacy Scale (Galesic & Garcia-Retamero, 2011), to more objectively and comprehensively assess individual differences in chart-related knowledge and skills. The concept of data visualization literacy, built on reading and construction of data visualizations (Börner et al., 2019; S. Pandey & Ottley, 2023), has lately emerged in the field and could also offer a promising direction for future research.
Furthermore, we only included static charts, although dynamic charts have been increasingly present, at least since Gapminder (Rosling, 2007) stellar success. However, dynamic charts are posing new research questions (Sutherland & Ridgway, 2017) that go beyond our current study aims. Emerging technologies, such as AI and interactive visualization platforms, offer exciting opportunities to personalize and adapt chart designs to different user needs, potentially transforming how data visualizations communicate complex information.
Despite these limitations, this study provides valuable insights about how comprehension and recall can be examined using authentic materials (e.g., Eurostat charts) in ecologically valid settings.
Conclusion
In an era increasingly shaped by data, improving how visualizations communicate information is vital for combating misinformation and empowering individuals to make informed decisions. This study examined the factors associated with the comprehension and recall of bar charts, finding that effective data visualization requires both enhanced chart literacy (user side) and evidence-based design improvements (provider side).
Our study affirms that while well-known bar charts are easily grasped at a surface level, deeper comprehension—identifying patterns and drawing inferences—is less common. This points to a need for educational interventions that move beyond basic reading to foster genuine critical engagement with visual data. The positive correlation we found between performance and participant background (experience with statistics and charts) suggests that visualization comprehension is a teachable skill, not a fixed trait. This necessitates greater investment in capacity building across all educational levels to improve statistical and data visualization literacy.
For chart design, our study finds that embellishing bar charts did not harm messaging and significantly aided recall. This suggests that strategic design elements can be used to enhance memory and engagement. Effective chart comprehension is thus shaped by a combination of the data, the design choices, and the users’ prior knowledge, and context.
Ultimately, improving data visualization literacy requires a holistic approach: better design practices, richer learning environments, and targeted training. As the ability to understand and communicate with visual data becomes increasingly essential, our findings offer practical guidance for educators and designers alike, calling for renewed attention to how we teach people to see meaning in data. Hence, we recommend further exploration of ways to achieve effective and sustainable chart messaging, focusing on the interplay between user background, design enhancements, and cognitive outcomes.
Footnotes
Acknowledgements
We thank the anonymous reviewers for their helpful comments and all respondents for their participation. Our gratitude goes to Jerneja Kos for technical support and testing, and to the mentors and doctoral committee for their guidance. This work was partially funded by the Slovenian Research Agency (No. P5-0441). Eurostat charts are licensed under the Creative Commons Attribution 4.0 International license.
Ethical Considerations
This study involved an online study on the Prolific research platform that collected neither personally identifiable nor other sensitive information, so formal ethical approval was not sought. The study was conducted in strict adherence to ethical best practices.
Consent to Participate
Participants were fully informed about the study’s purpose, type of questions, use of collected data, contacts and their rights, and they explicitly provided informed consent before proceeding with the survey.
Funding
The authors disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: The work was partially funded by the Slovenian Research Agency (No. P5-0441).
Declaration of Conflicting Interests
The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.