Health Data Visualization Literacy Skills of Young Adults with Down Syndrome and the Barriers to Inference-making

One foundational skill often used in health data visualizations is the ability to read. Reading involves constructing understanding of a shared meaning between a writing system (i.e., print) and its corresponding verbal or gestural language. Eighty-six percent of the world knows how to read and write [121]. However, between 4.9% to 27.7% (mean: 15.5%) of the 86% of literate adults (16–65 years old) have poor reading proficiency [33]. The total number of adults with poor reading proficiency could be higher as individuals with “language difficulties, or learning or mental disabilities” were excluded from this research effort [33].

Print literacy is comprised of multiple, complex language skills that underpin a person’s reading abilities. These underpinning skills and abilities are: phonology (i.e., speech sounds), orthography (i.e., spelling patterns), semantics (i.e., connecting symbols of written language to their meaning), and morphology (i.e., patterns found within word formation and language syntax) [88]. A further set of discrete decoding skills are necessary to move individuals from responding to visual and auditory stimuli in a written or spoken language towards competency (i.e., accurately and effectively constructing both meaning and their understanding): (1) Phonemic awareness is the ability to recognize and use with the individual sounds in spoken words; (2) phonics matches sounds to their corresponding letters or groups of letters; (3) fluency is the ability to read accurately and quickly; (4) vocabulary involves increasing the known words used in a particular language; and finally, (5) comprehension occurs when reading words constructs understanding [88].

An individual’s difficulty or ease in which they can grow their print literacy skills are closely related not only to adequate instruction and regular practice, but also to the body. For example, an individual’s hearing is involved during phonics activities in verbal languages; their vision is used to recognize orthographic elements, letters, letter groups, words, and phrases; and their working-memory is used to connect semantic meanings, recall vocabulary, comprehend the content, and construct their understanding of what they just read [88]. However, vision, hearing, and working-memory can be impacted by developmental differences within individuals with Down Syndrome [14]. As a result, reading abilities may vary across individuals.

Despite an increased potential for the kinds of impairments that may impact someone’s print literacy abilities, researchers have found that decoding words (i.e., figuring out a new word by sounding it out) and single word reading are relative strengths [14, 87]. However, people with Down Syndrome can struggle with phonological awareness [76]. Reading interventions that simultaneously develop their speech and phonological abilities whilst also growing their vocabulary and grammar skills may increase their abilities [58, 87]. Multi-modal interactions that reinforce the sound-word mental connections may be beneficial in digital health data visualizations. Given the unfamiliarity of many medical terms used in health systems and data visualizations, ensuring adequate vocabulary supports are incorporated may further facilitate the reading abilities of individuals with Down Syndrome.

Roughly one-third of English-speaking Americans (\(\sim\)62.7 million) have low numeracy skills [80]. Numeracy is someone’s ability to use, reason with, and understand math in their daily lives [1]. Numeracy skills include arithmetic and mathematical skills, such as counting, understanding number lines, numerical concepts (e.g., whole numbers, percentages), number sense (i.e., the cognitive flexibility to take numbers apart and put them back together), performing calculations (i.e., addition, subtraction, multiplication, and division), operation sense, measurement concepts and protocols, estimation, proportions, and more [44]. As such, numeracy is also essential to understanding data visualizations [1]. Health numeracy is “the degree to which individuals have the capacity to access, process, interpret, communicate, and act on numerical, quantitative, graphical, biostatistical, and probabilistic health information needed to make effective health decisions” [49]. Sufficient health numeracy skills are critical to understanding risk, probability, and scientific evidence (i.e., medical information) [111].

As data visualizations are often intended to support human comprehension, pattern finding, and inference-making of numerical information, an individual who struggles with health numeracy and graphicacy may encounter unnecessary barriers to their ability to make effective, informed decisions [15, 120]. Unfortunately, people with poor numeracy skills are also more likely to have poorer graphicacy skills [45]. Similarly, people who struggle with numeracy skills, such as those with Down Syndrome, may also experience difficulty with abstract reasoning and other essential skills [35, 40, 74, 83] that can impact their data visualization literacy skills.

Data literacy is the newest type of essential 21^st century skill necessary for full participation in an increasingly data-driven society [28]. Yet despite being a critical skill, there is no widely accepted definition for data literacy [130]. Some definitions broadly describe data literacy in terms of using and understanding data (e.g., [34, 47, 65, 100, 103, 105, 108]) while others see it as an inquiry process where hypotheses are formed or questions are generated about the data to either solve a problem or support decision-making (e.g., [36, 53, 79, 95, 131]).

Data usage can include multiple sub-processes relating to reading and manipulating data, such as connecting, comparing, distinguishing between data, transforming data into information, analyzing, data processing, and communicating conclusions [66, 69, 78, 79, 100, 107, 112, 117, 129, 135]. Print literacy, numeracy, and other statistical knowledge are requisite underlying skills to be data literate [5, 28, 50, 135].

Critical thinking skills are also used throughout the process to support an individual’s analysis, inference-making, and ability to synthesize the data to reach a conclusion or to explore potential solutions [65, 106]. However, some critical analysis skills fall outside of this study’s scope of investigation, such as the critical evaluation of data accuracy via viewer assessment of the methodologies used to collect the data and potential biases of the individual’s personal interpretation [28, 59, 107].

Data literacy is often described as an ability [53, 78, 79, 95, 100, 103, 117, 122, 129]. As several scholars note that an essential skill of a data literate person is their ability to read data visualizations [5, 16, 23, 78, 95, 117], ability-oriented definitions suggest that when someone is hindered, such as by inaccessible data presentations, an individual’s potential to leverage their personal data is limited.

Roughly 29% of Americans have poor data visualization literacy (i.e., graphicacy) [94]. Graphicacy skills can help viewers understand data visualizations. Graphicacy involves three stages of activities that work towards completing specific goals. The earliest, foundational phase has the goal of reading the visually represented data; the second phase’s goal builds up the initial information by connecting and mapping the information together within the visualization; and the third phase uses the information from the earlier phases to support the viewer’s ability to extract meaning and make inferences about the content presented with information outside of the visualizations [17, 73]. Curcio broadly described these phases as: reading the data, reading between the data, and reading beyond the data [29].

The first stage—reading the health data—involves various sense-making processes (i.e., taking in and interpreting visual information) to identify the various component parts that make up the visualization as a whole. These sense-making sub-activities are necessary to reach the first stage goal of reading the health data. Examples of sense-making activities include the viewer identifying the elements that make up the health data visualization’s anatomy (e.g., title, X- & Y axes, labels, values, numbers, special characters, symbols), the value encoded elements (e.g., shape, size, position), and the type of data visualization, which can influence how the viewer might approach the graph reading task. This initial visual encoding of information helps people to determine what is salient–and, thus, worth their attention and cognitive efforts to process [89].

The second stage–reading between the health data–involves comparing and connecting information to identify patterns, differences, and outliers. During this stage, people use various information foraging sub-activities [97] and other sense-making processes [17] to reach this phase’s goal of tying together how the graphical information is connected with each other and what these connections may mean. This stage involves an iterative process of searching through the data, visually encoding it, and then mapping the information presented to the viewer’s internal mental model of the visualization [77]. Reading between the data tasks can include things like identifying extremes, finding exact values, anomalies, or clusters, making estimates, noting ranges of values, connecting graph elements or other information to the data (e.g., relating the x- & y-axis information to the graph values, connecting meaning to stylistic elements, like color), and comparing any differences in the data that may occur over time (i.e., noticing trends and recurring behaviors).

In the final stage, viewers draw more heavily upon information held in their long-term memory (e.g., health knowledge and other related literacies or skills) to inform their understanding, perform inference-making, and update their mental model as they engage with the data visualization [113]. During this stage, the information is especially influenced by any domain-specific knowledge they possess. This is used to extract the information in an informed way to read beyond the data. It is during this stage that individuals leverage the information in the visualization to reflect upon their health behaviors and gain potential insights for where improvements to their health could be made. All three stages, but especially the final stage, are driven by knowledge. It is also important to note that these stages of reading a data visualization are not necessarily a linear process. Rather, viewers may move back and forth between stages as they refine and update their mental model as they take in more information and make new connections using information from both within the visualization and with information outside of it.

Health literacy involves multiple processes related to finding relevant information, understanding the information that is being accessed, and critically evaluating health information (i.e., interpret, filter, and judge the information) to make informed healthcare decisions and communicate those decisions with those involved in the management of one’s health [9, 19, 118]. Health literacy skills require print literacy, numeracy, and oral literacy when the information is shared with others (i.e., members of the medical care team and trusted adults involved in the shared or supported-decision making process).

Data visualization literacy-based skills within the medical context also requires health literacy knowledge and skills to inform the viewer how to read, interpret, and engage with the graphs or charts. People must use various information-seeking skills, understand new health information, and synthesize it while interpreting their health data, as described by information foraging theory [48, 97, 124]. A viewer’s ability to search, interpret, and make sense of any health information they find is hindered if they lack the requisite health literacy skills and domain knowledge [124, 127].

Nutbeam identified three levels of health literacy: functional (i.e., ability to understand basic health information within the healthcare context), interactive (i.e., engages, asks questions, applies new information to changing circumstances, and incorporates that knowledge into shared decision-making with healthcare professionals and/or family), and critical health literacy (i.e., critically analyze health information to make an informed decision that aligns with personal preferences and needs, including when to get a second opinion) [90]. Unfortunately, many people have poor health literacy [3, 35, 110]. More than one-third of Americans have insufficient health literacy skills; only 12% were proficient enough to effectively use health information [92].

The foundation of reading any graph or chart begins with identifying the various elements that make up the health data visualization. We began our interviews by asking participants some open-end saliency questions, such as “What is the first thing you see?” During the second saliency question, the researcher physically covered their own eyes with their hands and said “I am closing my eyes now so I can’t see.” The researcher then asked participants to describe everything they were seeing in the graph to them. The benefit of taking a saliency-first approach in the interview was two-fold. First, the saliency questions show us generally where the participant was looking and in what order they read the graph elements (i.e., how they naturally read graphs without any kind of structure). Secondly, the act of verbalizing provided insight into the saliency judgments they were making (e.g., an element’s relevance, importance, noticeability, etc.). This combination demonstrated how people with Down Syndrome initially read health graphs without any supports or guidance.

Participants described the HDVs in four different ways (see Appendix A.3 for all saliency HDVs). First, they made generalizations based on their initial assessments of the information presented, compared different types of data or regions in the HDV, categorized types of data, and made judgments about the HDV. For example, upon seeing the different images of food in the scatterplot, most participants generalized that the data was split into two groups, healthy and unhealthy foods and beverages. Participants also called out specific types of graph elements. These could include the type of visualization (i.e., bar or line graph), the use of icons or images, the presence of words, numbers or dates (e.g., “I see a lot of numbers” in the table). Participants also vocalized elements’ specific text content, such as the title in verbatim or an actual number value (e.g., “2.0”) in the HDV. The last type of observation that our participants made was noting the various descriptive qualities (e.g., the color, size, or position) of HDV elements.

All participants made multiple combinations of descriptions for every HDV. This demonstrated that every participant was capable of varying levels of abstract thinking upon seeing an HDV for the first time. The number of observations verbalized also differed between participants. The quantity and locations of observations similarly indicate different graph reading patterns across our participants. It also showed the visual path they were taking, where they visually focused on a region or if they returned to a region or element more than once. As such, the order of the observations likewise suggests varying levels of ability in effective scanning of information.

When initially reviewing the HDV, participants often leaned upon skills they were strongest in. For example, most participants employed their print literacy skills first. Many participants demonstrated a tendency to read specific text elements first. Participants then read text usually from the largest to smallest in text size. This was followed by participants verbalizing items that had sufficient color contrast as this facilitated quick and easy recognition of information. After the written text, participants demonstrated a tendency to either call out familiar icons or images or specific colors when present in the visualization. Next, participants verbalized numerical information. The order of numerical elements that were verbalized similarly followed the largest to smallest with strongest to weakest color contrast. This suggests that familiarity and confidence in various skills may influence what is most salient and in what order this population reads HDVs. In other words, many individuals were immediately drawn to HDV elements they were confident in their ability to make sense of the more familiar elements (e.g., print, numbers). This familiarity-first behavior suggests some viewers with Down Syndrome may employ a combination of top-down (i.e., long-term knowledge and skills) and bottom-up (i.e., taking in stimuli without context) information processing as they interpreted the information to construct their understanding of an HDV at the earliest stage of graph interaction.

As people with Down Syndrome can struggle to express themselves, it is worth noting that several participants frequently avoid saying words or numbers that they struggled to audibly say. For example, many participants struggled with saying the word “macronutrient” in the stacked bar chart. Instead, they talked around, were hesitant to say, or omitted the word (e.g., “Whatever that long word is called” [Harper], “It’s a bit hard to say” [Morgan] “This week’s–not–uh, I–I don’t know” [Jordan]). However, participants struggled less when they broke down the syllables. When Emery got to graph #3, their study partner covered up the syllables as they read them, allowing their to gradually read the word. This may signal an accessible design opportunity for long words or jargon in HDVs with populations who may struggle to visually parse through multisyllabic words (e.g., “This week’s Mac-ro-nu-tri-ents”).

Similarly, verbalizing numerical information also highlighted the numeracy issues people with Down Syndrome can experience. There appeared to be an issue when reading long number place values. Past work has suggested that dyscalculia is often a part of the Down Syndrome behavioral phenotype [30], which occurred when participants encountered extra characters, such as decimals separating the whole and fractional numbers, commas indicating higher-level number place values (i.e., tens, hundreds, thousands), or dashes used in ranges. For example, Emery read the age range “31–35” as “135” in the first table. When Jesse said, “18 and 20, and 20 and 39, and 62, 85,” they were actually reading the age range (18–20), the first two cells (2.0, 3.9) in the blue Lean region and first two cells (6.2, 8.5) in the green Ideal table region. Similarly, Cameron read the total value of 63,451 in the bar chart as “Six thousand–six, five, thirty-four, fifty-one.” Skyler had to correct themselves when reading the average number of steps “9,600–9,064.” This may indicate incorrect number articulation errors caused by special characters creating visual shifts as they read the number and encoded the position with the appropriate number place values.

Invisible number lines could also be a stumbling block for some participants. Jesse similarly could not find their age: “I’m 22. but 22’s not on here.” According to the table, Jesse would fall into the 21–25 age bracket. In this example, the use of age ranges requires a viewer to have both sufficient working memory and numeracy skills to recognize the invisible number line with a range of numbers. Increments on the X- and Y-Axes may also introduce invisible number line barriers. On the “Weekly Walk Distance History” line graph, Darcy observed: “I see that the graph skipped some numbers. It starts from 2.0 and then it is counting by five. Five, zero, five.” Sloane struggled to articulate their frustration with the increments–in the y-axis of the macronutrients stacked bar chart. They referred to these skipped numbers as categories: “I see 100 category. I see two–2,000 in categories and three cat–category and forty catty–category, ach!”

An essential part of reading a health data visualization is the ability to identify the various graph component elements (i.e., title, axes labels and values, icons, images). HDV element identification requires sense-making, spatial awareness, and effective scanning abilities. Additionally, interaction with a data visualization is not a linear process [17]. Instead, someone’s interaction and understanding of an HDV is continually refined as more information is visually sensed and encoded and mental models for the data are iteratively revised and updated. Table 2 reports upon participant performance of each sub-task within this first stage of HDV reading. Correct answers are indicated by a point and incorrect answers with zero points for each graph and each corresponding question.

Overall, the participants performed well when identifying graph elements such as the: title, X-axis labels and values, Y-axis labels and values, values within the visualization, and various symbols, icons, or images. They also were able to identify the various stylistic elements such as shape, color, and size. Six of the ten participants were able to identify more than 75% of the various HDV component parts. The remaining four were able to identify between 65.5% to 72.4% of the elements. This suggests that the graph perceptual, sense-making skills of people with Down Syndrome are relatively strong during early health data visualization reading identification activities. This is in line with past work that found people with Down Syndrome to be strong visual learners [41].

4.1.2.1  The X- and Y-Axes. The X- and Y-axes labels scored the lowest during graph element identification. This appeared to occur, in part, because of a behavioral reaction to pictures when influenced strongly by existing health knowledge particularly in the scatterplot graph (graph #6). The scatter plot’s use of strong, realistic pictures appeared to reinforce our participant’s existing nutritional understanding. As a result, the X- and Y-axes labels were often ignored. This visual disregard may have occurred because photographs of food were used. Viewers appeared to instead fixate upon the highly familiar, concrete data points rather than noticing what was being compared on the scatter plot, which compared the percentage of Americans versus Nutritionists who said whether a food or drink was healthy or not.

Participants demonstrated an observable tendency to first notice the X-axis, followed by the Y-axis on the left whenever participants read graphs. However, many participants (60%) struggled to notice an additional Y-axis was included in graph 5. Only three participants noticed both the label and the percentage values [Harper, Skyler, and Sloane] when a dual Y-axis was present. This behavior may indicate that some participants were unfamiliar with the procedure when reading different types of graphs with axes in different quantities and locations.

4.1.2.2  Icons vs. Images. In the HDVs that used icons (graph # 2–5), participants verbalized more descriptive qualities (e.g., colors, or data types, such as numbers or words) when they generalized what they were seeing in addition to a greater number of specific HDV details as they took in the information. Conversely, most participants (80%) in the scatter plot HDV used more generalized descriptions of large categories of information. It appeared that the use of images of various food and drinks combined with an unfamiliar graph type caused them to rely more upon their understanding of nutrition. This resulted in broad generalizations about the data points and the visualization as a whole that was directly informed by their pre-existing nutritional knowledge. Participants also categorized and grouped the more familiar imagery (e.g., foods and drinks, healthy and unhealthy, “fat stuff” and “too much salty” [Morgan]). This kind of data categorization and grouping can support estimation abilities [119]. Participants, like Cameron, made other associations with what they were seeing as well: “I see different types of ... foods in the kitchen, oven and stuff and the cups.”

Most participants (80%) skipped a lot of individual data point identification that they had demonstrated in previous HDVs. Instead, many went straight to making judgments about what was being depicted. This interestingly suggests that mental models of HDVs could become less flexible if pictures are used on their own without additional elements that can support accurate inference-making if an assumption is made based solely on the familiarity of the images in conjunction with their initial impressions of the data. It appeared that the confidence experienced by the participants caused by the combination of familiar presentation and topic that simultaneously reinforced their existing nutritional knowledge made them less visually critical of the remaining HDV elements and what the relationship between data points could mean, especially if it conflicted with their existing understanding of the topic being visualized.

Mapping information in HDVs generally consists of connecting the identified component elements of the graph’s anatomy to each other and encoding the meaning of visual attributes. Mapping information in this way supports viewers’ ability to assign meaning to each connected element and update their overall understanding of the visualization. In this subsection, we describe the two visual attributes these HDV employed to support interpretation of categorical information encoding: color and images.

4.2.1.1  Color Encoding and Meaning Mapping. Color is typically used to distinguish categorical information by grouping elements so viewers can more easily identify similarities and differences. Many systems designed for people with Down Syndrome or other IDDs, use color-coding to indicate more than just groups. Some nutrition-oriented health apps use the color metaphor of a stop light to indicate a food’s health status (e.g., [71, 101]. However, graphical properties, like color, are not equal in their ability to accurately communicate meaning. Other channels, such as spatial region, position, length, angle, and size, are more effective [25, 84].

Three graphs used color encoding: the table (#1), the stacked bar (#3), and the line graph with two y-axes (#5). The table used both colors and labels to associate the two as a group (i.e., blue = lean, green = ideal, yellow = average, red = above average body fat percentage). In the line graph, color coding was used to indicate intensity of a physical activity to visually link the heart rate and effort y-axes together. Finally, the stacked bar used color to indicate the healthiness of a macronutrient using the stoplight visual metaphor, to distinguish between the three values, and to visually link the macronutrient label with the icon (i.e., lettuce = healthy carbs, leg of meat = protein, butter = fats).

When color-coding did occur, 30% of the participants had no meaning association. Instead, viewers, like Jesse, inferred that color was simply a stylistic choice: “it means the colors ... like different kinds of colors. It’s blue, green, yellow, red.” The percentage of those who had no meaning associated with color was higher during the first half of the interview versus after spending more time engaging with the HDV. Harper was the only one who explicitly stated that the graphs were “color-coded.”

In the table, many participants recalled a pre-existing color association, “Normally, the green, yellow, and red means ... Stop, Slow, Go. But [I’m] not sure about the blue” [Harper]. When other colors did not also map to the metaphor, like the blue, confusion occurred. For Jordan, green was “good,” yellow was “bad,” and red was “very bad.” However, they instead associated blue with the affective state of “sad,” which may indicate a color metaphor-mismatch occurred. This inaccurate encoding made interpreting the graph more difficult. Typically a high-performer, Cameron did not notice the labels at all. Instead they interpreted the color as corresponding with the size of the region. While Sloane did connect blue with its lean label, they said green was “healthy” and red was “really bad.” They described yellow in terms of foods that were both healthy and yellow-colored.

Graph #5 similarly indicated how more explicit mapping between color and meaning is necessary in more accessible HDVs viewed by people with Down Syndrome. High-performers were able to associate how color was used. Skyler observed that colors were the “different kind of colors of different beats in your heart.” Similarly, Cameron was able to connect the title and associate the color with its gradation: “The colors mean how–how deep is the intensity... [Green] means like not that–not that intense. Dark red means that it’s that extreme amount in the intensity.” However, imperfect color encoding and mapping still occurred 45% of the time. Jordan associated the color with the icons rather than intensity levels. This suggests that despite the presence of labels, color is not strong enough to ensure accurate mental connections in a world of potential implied meanings.

The stacked bar chart had the highest level of correct color mapping at 95% accuracy. Even participants, who consistently struggled [Shiloh, Emery, and Harper], were able to connect both color and icon encoding when properly reinforced with familiar distinct imagery that was reinforced by understanding of health information and had a label to support mapping. However Jordan’s concrete associations with the food icons overruled a totally accurate color interpretation: “[Red is] bad food. [Yellow is] good food. [And green is my] favorite food.” This may signal the strength of lived experience to inform HDV interpretation.

4.2.1.2  Image Encoding and Meaning Mapping. Five of the six HDVs used icons or pictures to support comprehension as suggested by previous accessible visualization recommendations [134]. While icons did support most participants’ connection between graph elements, some participants did not like the use of icons. Darcy, who self-reported that they were very familiar with reading graphs, preferred a “dot” instead of the shoe icon. As such, there are some caveats for how those icons could be incorporated into visualizations intended to be more accessible to people with Down Syndrome within the health context.

Appropriate image selection is critical. Sloane initially thought the vertically stacked shoe icons in the Daily Steps bar graph was a “shoe store.” While Cameron extrapolated that “those lo–logos represents [the] amount of di–distance” in the Weekly Walk Distance line graph. Others, like Jordan, associated the shoe icons not with steps but with physical activity (i.e., “walking”) in the bar graph. However, for Darcy, the shoes were simply shoes and the caution sign meant nothing to Morgan in the dual Y-axis line graph. Sloane interpreted the lettuce icon in the stacked bar chart as green brain, which, in turn, impacted their connection of the icon to its healthy carbohydrates label.

Graph complexity can also interfere with accurate icon encoding. The use of multiple icons in a single graph should be carefully considered, particularly if there is not a label to provide a redundant encoding for bundling connections between elements. In the fifth graph, while the use of different icons were intended to visually represent a change in activity intensity to reinforce position on the graph as well as the color encoding on the graph background, the different icons did not clearly map. As a result, the complexity and use of multiple icons in the dual y-axis line graph contributed to the icons only being 45% accurately encoded–the lowest scoring graph.

Icon encoding can take time to process. Like some, Morgan did not initially verbally associate the shoe icons with steps. However after spending time answering questions about the bar graph, those participants persistently associated the shoe icons with steps across all three graphs that used the shoe icon. This was interesting as the meaning of the shoe icon changed with every graph that used it (i.e., bar graph #2 = steps; line graph #4 = distance; dual y-axis line graph #5 = mid-level intensity). The re-use of icons when an association has been made caused participants, like Morgan and Harper, to consistently demonstrate this carryover effect during icon encoding. Other icons also had pre-existing associations, which similarly affected correct meaning-mapping. The use of hearts and a warning sign to indicate higher heart rate zones meant something different for Harper, Morgan, and Jordan, who interpreted the hearts as “love” or “falling in love” and the triangular warning symbol as “person” [Harper], “danger” [Shiloh and Sloane] or a “danger zone” [Darcy].

The results from this and the previous sections may signal an accessible HDV design feature opportunity to provide more explicit inference-making features that support the connection-making abilities of individuals connecting graph elements and channel encoding, like color or icons.

Comparing information in HDVs involves both visuo-spatial and abstract reasoning as well as ratio-processing abilities to recognize patterns, similarities, differences, extremes, anomalies, and ultimately connect the various elements together to infer the overall trend. This section describes how people with Down Syndrome compared high and low extreme values. It also reports upon the differences participants noticed when viewing HDVs. The number and kind of differences indicated varying levels of cognitive flexibility among our sample when recognizing patterns, generalizing, and grouping.

4.2.2.1  Comparing Extreme Values. The extreme value questions illustrated authentic reading tasks where the viewer begins with a specific question to find an exact value (see Tables 14 and 15 in Appendix A.2 for the exact HDV specific questions asked). Overall, participants performed moderately well when comparing data to determine extreme values. They were able to accurately discern both the high values roughly half of the time: Table (60%), Bar (50%), general values Stacked Bar (50%) specific values Stacked Bar (60%), Line (45%), Dual Y Line (45%), and scatter plot (45%). They performed slightly better when judging low values: Table (85%), Bar (70%), general values Stacked Bar (50%), specific values Stacked Bar (60%), Line (50%), Dual Y Line (55%), and scatter plot (40%). One reason for the moderate performance may be that participants had to use different procedural skills to locate extreme values across all six HDVs.

For example, the table (HDV #1) required effective scanning of a large amount of numbers, some with poor color contrast. While Shiloh generalized to entire regions, some participants, such as Morgan and Sloane, instead visually fixated and answered within their field of vision when the question was asked. The table also required participants to recognize patterns across the overall numbers to find the location of extremes. The values increased from top-to-bottom and left-to-right with the lowest in the upper-left and the highest in the bottom-right. There were also table reading procedural issues for Jesse and Sloane, who answered with age values, which were literally the highest number visible to them.

HDVs #2-5 required participants to visually track between the graph’s axes and the individual values. During this visual back-and-forth, participants engaged their ratio-processing system to visually compare the differences between values. The brain’s ratio-processing system attends to ratios of difference between non-symbolic values (i.e., not number symbols, but shapes or areas) [82]. As HDVs are visualizations of non-symbolic values, comparing between them requires noticing fractional differences between the information represented. When the contrast between values is great, it requires less cognitive effort. When it is smaller, it can increase a viewer’s cognitive load (Figure 6). After this comparison has occurred, they then must keep track of each of the value judgments in their visuo-spatial working memory until they find their answer.

Comparing the lowest values in the bar graph required the least cognitive effort for ratio-processing. This made sense as most participants were able to correctly answer this question. However, more sensitive ratio-processing skills were necessary for the high values in the bar graph and both the overall high and low extremes as well as the specific macronutrient type extremes in the stacked bar graph. Half of the participants struggled to correctly determine the highest value in the Daily Steps bar chart. Saturday (i.e., the correct answer), Tuesday, Wednesday, and Thursday all had very similar values to each other. Because of this lower contrast ratio, four out of the five answered one of the visually similar, yet incorrect values (Emery, Harper, Darcy, and Jordan).

Ratio-processing skills were taxed in the stacked bar graph when participants were asked to determine extremes of specific macronutrient types (i.e., highest protein and lowest fat). Unlike comparing values in the bar graph, which started on the same level on the x-axis, comparing fats and protein areas were more difficult because they sat on top of the different healthy carbs values. Being stacked on an uneven base appeared to impact participant’s abilities to effectively process and differentiate between the lower contrast, fractional differences in the sizes of the protein and fat rectangles. This suggests an accessible visualization design opportunity to highlight when the values have low ratio contrast and reduce unnecessary impacts to the viewer’s cognitive load.

When participants had hit their ratio-processing limit, they reverted to a more familiar graph reading strategy of looking to the top of the shape to determine the highest or lowest values. All four of the participants (Emery, Harper, Darcy, and Cameron), who incorrectly answered Thursday as the day with the most fats, did so because Thursday had the most overall grams. Visually, Thursday had the highest position from the top. Similarly, when Shiloh was uncertain how to judge the highest and lowest values in the daily steps bar graph, they reverted back to their stronger number reading skills as they seemed less confident in the ratio-processing abilities. In both instances, Shiloh answered with the smaller numerical metric underneath the title: “Yeah, so the highest and the lowest: 63,451. And the lowest is the nine thousand. No, no. Nine hundred, sixty-four ... so–the highest and the lowest.”

While the imagery was the most familiar, the procedural skills required to read a scatterplot graph was the most foreign to all participants. Several participants leveraged other graphicacy skills they felt more confident using as they interpreted the nutritional data. For example, two participants reported that kale [Darcy and Sloane], which was visually in the highest position to the top of the graph, was the most healthy. Other participants relied upon their health literacy and were instead influenced directly by their nutritional knowledge. In the scatter plot, participants' broad generalizations were impacted by their existing understanding of nutrition: “the highest food is the healthy foods” [Skyler], “the veggies up top” [Emery] or “the fruit and vegetables” [Harper]. Shiloh said the “junk. It’s ice cream–the desserts” were the least healthy in the scatterplot. Harper echoed this assessment by generalizing with “all the junk.” As a result, how participants compared values suggests that when individuals with Down Syndrome felt less confident in their interpretive procedural skills they instead switched to more familiar skill sets and prior knowledge to interpret HDVs. In other words, they use the same tactics as other typically developing populations when they are not sure how best to proceed when they are uncertain how to interpret a data visualization.

4.2.2.2  Comparing Differences. The ability to effectively compare values within an HDV is critical to notice patterns within the data, where values diverge from each other, and what anomalies or outliers there may be. Effective comparative skills support the viewer’s understanding of the data by examining the relationships between values. Comparing the relationships between and across the HDV is foundational to interpreting patterns on the micro (e.g., [in]consistency of a performance instance) and macro level (e.g., overall trend across multiple instances).

All of the participants were able to compare data and demonstrated varying levels of cognitive flexibility when answering this question. Skyler showed the highest level of flexibility when mentioning differences in HDVs–a total of 18 differences across the six graphs. Cameron was the second highest at twelve and Emery reported eight. The lowest was Harper, who noticed three differences.

Participants verbalized multiple kinds of differences. The types of differences mentioned were: the individual shapes, differences between regions, overall trend across the visualizations, colors, and images. The most commonly reported difference type was the various kinds of data (i.e., text, numbers, axis labels and values, specific graph values). This once again highlights our sample’s tendency to rely upon and leverage their strongest skills (i.e., reading) when doing an unfamiliar task like verbalizing differences of data representations.

After an individual has encoded and mapped meaning between the elements of the HDV and has compared values to get the gist of the data, the viewer will then connect those component elements together. Connecting information allows viewers to make sense and begin to understand the overall nature of the HDV. Creating mental relationships between the information allows the viewer to understand: (1) the topic and (2) the overall trend.

Although most participants performed moderately well throughout the entire interview, performance dropped dramatically as many participants struggled to connect the various elements and aspects of an HDV to synthesize a coherent understanding of the information represented. Results support previous work which found that while people with Down Syndrome understand abstract information, the differences to their working memory can make managing too much information with too many relationships at the same time a challenge [20, 60].

4.2.3.1  Identifying the Topic. Only a quarter of the participants were able to connect the various graph elements, aspects, and information to the overall topic of the table (HDV #1). A little less than half (45%) of the participants were able to get the overarching topic of the bar chart, the stacked bar, and the dual Y-axis line graph. The HDVs with the highest levels of connection between the data and the topic were the line graph and the scatterplot at 50%. It is worth noting that unfamiliarity with the scatterplot graph type led to everyone generalizing. Everyone got partial credit for verbalizing that the graph depicted the overall topic of healthy and unhealthy foods and drinks. However, no one was able to provide the more nuanced answer: the scatterplot was comparing the perceptions of healthiness of food and drink items judged by nutritionists versus the average American.

Cameron, Darcy, Emery, and Harper were the most consistent individuals to succinctly synthesize and summarize many of the HDVs. Harper described the table being: “about the ages and the percent of the fat.” Emery connected the HDV to their everyday life, which made connecting the data to the topic much easier: “It is called steps for–same as my watch! It tells you, like, activities in there.” Cameron recognized that the macronutrient stacked bar graph was about: “the grams of, like, the amount of food. And the food has different categories because carbs, protein, and fats.” Emery described the line graph as: “It looks like a snake. That is how many walks have you done. ... how much you’ve done it–of the walk distance history.” Cameron summarized the topic of the dual Y-axis line graph by saying: “it’s all about the intensity in the activity, um, it tells you ... the times at the bottom. Um, is telling you about different times of the levels of the activity intensity.”

A mixture of partially correct and incorrect answers indicated that participants were influenced by their personal understanding of health, exercise, and nutritional knowledge. For example, Skyler said the table was about: “how much you eat ... It has different kind of colors of what–what the–the healthiest things that you can eat. That’s it in my head.” Like Jesse, Darcy described the Body Fat table as: “different pounds of weight you have and also about your losing.” Their response was informed by the visual similarities between the much more familiar BMI chart often seen in doctor’s offices. Emery recalled the food pyramid when they saw the stacked macronutrients bar graph: “it’s ... like, um, a food triangle one that is ... And then there’s a different one. Different one is, like, that equals healthy one, protein, and fat.”

When participants answered incorrectly, most responses consisted of describing and identifying elements and aspects of the HDV rather than connecting everything together. For example, Sloane described the stacked bar as: “combine as healthy. And it will really combine to protein and fat” and counted the number of data points in the dual y-axis line graph, which had “22. There’s 22 times. It’s something measuring from, uh, to 80. Uh, maybe 75. ... all about, uh, the line of, um, gray mark.”

The issue of multiple blocks of easier-to-read text came up during the topic question as well. For some participants, seeing these text blocks did cause some incorrect connections. “It’s graph about the weekend because it started on, uh, every month. Like, um, Jan–January, Feb–February, March, April, and May. Hmm... new year. I don’t know” [Sloane]. “It’s the steps been taken in the years—since the numbers” [Harper].

4.2.3.2  Identifying the Overall Trend. Trend identification was the most difficult question for our participants. Most participants struggled to connect how each of the individual data points worked together to describe the overall trend (i.e., how the parts describe the whole).

Participants answered correctly the most often when the trend was obvious. Participants performed the best in the Weekly Walking Distance History line graph at 40% correct identification of the upwards trend. Several participants actually described the overall trend as they were connecting the graph elements to the topic. “This graph is describing that the walk distance is increasing” [Darcy]. Cameron described the red line as: “a snake goes up. ... all the logos in the snake that–that goes up th–those lo–logos represents amount of d–distance.” When Sloane couldn’t find the words, they instead vocalized the change: “the highlighted [line] And it goes “whoooop.” ... the number [is] bigger.” They audibly changed pitch of the vowels from lower to higher as they said “whoooop” to express the changes in the increasing trend. Sloane again embodied their response to describe the positive trend in the scatterplot as well. Cameron drew upon their health and data literacy skills to demonstrate their understanding of how nutritional components of the foods depicted affected where they fall on the plot: “The nutrients number could change because of the sugar weight.”

Trend identification was particularly challenging in the table and the scatterplot. This is because trend identification in the table required observing the changes using the numerical information only and none of our participants could recall interacting with a scatterplot before, so unfamiliarity of the graph type made describing its trend difficult. It was also challenging when there was no discernible pattern (i.e., not clearly ascending, descending, or remaining around the same amount), which occurred in the bar graph. Of those who were able to correctly identify the trend, only Cameron could describe the trend in the table and the scatterplot. Darcy was the only participant who partially described the bar graph’s trend and two others were able to identify the stacked bar and dual y-axis line graph’s trends.

There appeared to be a relationship between participants who were detailed describers during the saliency questions and those who were able to describe the HDV’s trend. Many participants used more concrete language to describe how the visual changes in data points appeared across the entire visualization. Skyler, who excelled at descriptions throughout their interviews described the stacked bar’s trend as: “I know for a fact it’s wavy.” Oftentimes relating the abstract patterns to more familiar, concrete imagery made trend interpretation easier to articulate. Shiloh similarly described the dual y-axis line graph’s trend: “it’s like a noodle.” In the same HDV, Skyler described the initial upward climb and subsequent dips as “very like up and down. It’s like a roller coaster. ... a pool. But it goes straight, but it has the little roller coaster baby pool to me.”

A well-known data visualization design mantra in HCI suggests that users want an “overview first, zoom and filter, then details-on-demand” as they engage more deeply with whatever information is visualized [115]. However, these best practices may not be as intuitive for people with Down Syndrome, who often did not think HDVs were interactive. 60% of participants thought that the HDVs were not something they could click on (table and bar: 75%, dual y-axis and scatterplot: 60%, line: 50% and stacked bar: 30%). Then, we asked what they thought would happen if they did interact with it. More participants thought of potential ideas for interaction (bar, dual y-axis, and scatterplot: 50%, table and line: 40%, stacked: 30%).

Many said they could not mentally envision or had “no clue” [Harper and Shiloh] what would happen if they chose to click on anything. The unknown outcome made even the adventurous computer user, Shiloh, who will “Click click click click. Gestures clicking across two computers. ... all the time,” irresolute. Instead, unfamiliarity with the task domain appeared to increase their hesitation and reduced their desire to perform even exploratory clicks. Jordan confirmed that any click or interaction with a HDV would be a “surprise” to them as they could not imagine what would happen. When Jesse did finally click on the stacked bar graph, they said: “Uh I broke it.” They immediately blamed themselves–rather than the HDV or technology–for the lack of response from their action. This may indicate those who were similarly hesitant or declined to interact may also share a strong internalized locus of control that could contribute to feelings of low self-efficacy during unfamiliar tasks [109] like engaging with HDVs.

Only Skyler interacted with every HDV with unprompted, adventurous exploratory interactions. Others just verbalized where they would click or tap. Participants said they expected interactivity on the axis values or labels [Emery, Shiloh], graph values, such as numbers, lines, or bar graphs [Cameron, Jesse, and Jordan], icons or images [Shiloh, Jordan], or “anywhere” in general [Harper, Emery]. While hesitant to click, Shiloh and Morgan did use Google to look up terms they were uncertain about, like “clickable.” Shiloh, in particular, regularly used the browser to find definitions throughout the entire interview.

Of the seven participants who did expect something to occur after they interacted with it, only two mentioned expecting details-on-demand for most of the visualizations [Darcy and Cameron]. They both expected to see the same, additional weight information on the body fat table cells, the total steps for each bar, the “carbs, proteins and fats” [Darcy] “by the amount of the grams” [Cameron] for each day’s macronutrients, both the number of miles and the distance in the line graph, and the nutritional information in the scatterplot. The two only diverged in the dual Y-axis. Darcy wanted each icon to reveal the “activity you’re doing” and Cameron instead expected to see “the number amount of the intensity.”

Some participants also expected the same type of outcome for each kind of interaction with the HDV. Emery wanted clicking on the HDV to either hide or close it except for the bar graph. For the Daily steps, Emery wanted it to function just like the same visualization they regularly interacted with on their Apple smartwatch. “It’s going to put it three ways. The first one is the activity thing. Second one is the serving thing. And then [the] third one is the rewarding, um, award thing.”

However, almost everyone had different expectations for what that interaction should be. Jesse wanted different types of interaction outcomes upon clicking the HDV, such as video content from YouTube in the table, “An app comes up ... Like social media” for both bar graphs and “my FitBit” for the stacked bar, and “a story or ... like a movie or something” in the line graph. Others expected an animation of more information to “... come right at you, ... get bigger” (Skyler) and “pop right up” (Shiloh). Sloane liked the idea of having videos or agents to provide additional information support. “Like a chat ... like, uh, people talking” about how fats can be “really unhealthy ... not healthy at all.” Like Sloane, Skyler also suggested gamified elements: “the shoes will come flying at you. ... I would typically duck under ... and you can eat the lollipops” [as a reward]. Sloane described how icons could use animation, which might provide additional layers of redundant meaning encoding to better support viewer interpretation and conceptually link the image to the activity being visualized. They suggested animated icons like the “walking of the shoe” and “a heart beating” to reinforce the connection with BPM.

Even though participants didn’t generate questions or have information-seeking triggers when interacting with the HDVs around three-quarters of the time, the remaining 25% generated rich types of questions and info-seeking activities. These were oriented around: (1) wanting more explicitly connected information together, (2) wanting additional encoding supports for icon meaning and value, (3) step-by-step guides when they were uncertain how to progress, (4) wanting to limit visual messiness, and (5) clearer definitions for unfamiliar words or abstract concepts. This section also describes the visual aversion we observed in some participants when they encountered intimidating or unfamiliar content in HDVs.

Emery had difficulty connecting the relationships between various types of data. Emery wanted these connections to be “better explained.” Shiloh was likewise uncertain how the age brackets and the regions were related and what those connections meant. Skyler also wanted to support connecting the relationships between various graph elements together (i.e., how the walk distance line graph values and weeks in the x axis were related). Emery confirmed that they wanted additional information and support interpreting values to better understand why the extremes happened and what other values meant in the stacked bar chart. Finally, Sloane wanted an explanation of the trend in the Weekly Walking line graph.

When the icons were used, Skyler wanted to know what the shoes represented numerically. “I wish it was explained better by how many shoes are on the measuring sticks.” If icons are meant to support comprehension, that needs to be explicitly explained in the design. However, if each icon equals a specific number of items, the total rather than an implied calculation needs to be included. In other words, a legend that says “1 shoe = 500 steps” would be less effective than having the total label say “On Monday, you walked 9,400 Steps.”

Finally, several participants wanted the HDVs to be more neatly ordered as visual messiness hindered their ability to make connections. For example, Skyler wondered why the color blocks in the table were different sizes. As oftentimes size is used to indicate quantity, some participants were confused when they noticed that the red above average block was visually larger than the blue lean region, which was, in turn, smaller than the green ideal and yellow average classifications. Emery, Cameron, and Darcy wanted neater categorizations of healthy and unhealthy items as they were visually overwhelmed by the data points being “jammed together” [Cameron] in the scatterplot. Darcy suggested filtering content and drill-down functionality to make it easier to view. “I wish they would organize the graph better. Yeah, because like now it is like messy ... so someone would click on carb, and then all the carbs would go into a category, and then there would be protein it would go there and then fats would go there into different groups–3 groups.”

Quick to look up information in the browser, Shiloh immediately Googled the word “intensity” when they encountered it in the dual-y axis line graph. However, they grew frustrated at times with the unfamiliarity and difficulty of the task domain (i.e., reading, interpreting and engaging with HDVs) when search engines did not provide helpful results. When Jesse encountered the unfamiliar words and concepts of “healthy carbs. [and] mac-uh-nutrients,” they wanted additional information to better understand what these items meant. Jesse also wanted definitions for unfamiliar health metrics: “I don’t know what beats per minute B–BPM means.” Skyler instead preferred to ask people in their life what things meant.

For other participants, however, they demonstrated a visual aversion to unfamiliar graph elements or aspects that felt either unnecessarily difficult or intimidating to parse: “I’m not clicking at the words” [Morgan]. Throughout the interview, many participants appeared to prioritize their cognitive effort and adjusted what they chose to attend to. As a result, more abstract or unknown elements became blind spots to be ignored as the viewer attended to problems they felt confident enough in their skills to solve. Another example of this behavior was when many participants fell back on their print and number reading skills when they were uncertain of the steps required to interpret and understand the abstract visualizations.

Participants varied in the depth of the level of reflection they were able to engage in with the HDVs. For example, Jordan connected the Body Fat table to being more conscientious of their food intake by “eating salad.” Morgan also found the line graph to be a motivation for them: “It is telling me to walk a little bit far, and like ‘keep going, going, going’ and like ‘take a break to drink lots of water’ [and then] ‘keep walking, walking again.”’ An athlete, Darcy, who already tracked health information in various aspects of their life, often found inspiration for other ways they could leverage their personal information to become even more physically fit with every visualization: “The [table] is telling me I need to start, I need to weigh myself.” In the bar graph, Darcy reflected that “maybe I can start tracking–measuring my steps, and see how many total steps that I’ve taken Sunday until Saturday.” And the line graph inspired Darcy to take it a step further by also recording “how much walk distance I walked.” The stacked bar chart told them “watch out for macronutrients ... like how much carb I’m eating. How much protein and fat I’m eating.” The visualizations also made Darcy want to tweak how they currently tracked their workouts: “this is telling to graph–to graph the time I’m exercising. When I’m tracking activities, I put the day that I’m exercising. I put the date and the type of the type of exercise I’m doing. But this graph is different, this graph shows you what time you are exercising ... I just put the date and then I put what I’m doing.”

The language used in HDVs may be generalized with more abstract terms (e.g., an exercise HDV titled: “This Week’s Activity”) to incorporate the wide range of things that the original design intended describe with that one word. However, participants indicated that HDVs intended to be read by them could benefit from replacing generalized language with more specific, concrete descriptors to better support comprehension. As such, participants demonstrated a type-token distinction preference when discussing increasingly abstract language appeared in HDVs. Types are abstracted descriptive concepts (e.g., a sphere); tokens are specific, concrete instantiated objects (e.g., that orange leather basketball on the ground) [128]. This preference became evident when participants indicated wanting clarification about the types of abstract information in an HDV. For example, when researchers asked type-oriented questions that used generalized, plain language to describe abstract content or concepts, such as trends, participants wanted more information. Responses that clarified meaning instead used more specific token language. When token-oriented language was used, participants answered questions with greater ease as the concrete specificity of the language enabled them to visually orient their attention to the specific content being discussed in the HDV.

In line with past research, we still found that people with Down Syndrome can “handle abstract things in a relatively easy way” despite the abstraction level preference for the more specific, concrete token distinctions over more abstracted types [83]. Rather, concrete specificity that drew the participants’ attention to tokens rather than types enabled them to better understand abstract content or the nature of type concepts. Future HDV design features that support this preference may be more beneficial and accessible to people with Down Syndrome.

After someone has identified the various elements of a health data visualization in the first stage, a viewer must also recognize if any information is missing that is necessary to effectively read the graph during later activities with the HDV. A fill-in-the-blank barrier is an umbrella term we used when some aspect of the data visualization design is implied, unexplicit, abbreviated, or missing in the HDV that forces the viewer to make a cognitive leap when attempting to make sense of the information. The barrier occurs due to a mismatch in ability to make the cognitive leap. To make a cognitive leap caused by a fill-in-the-blank barrier, a viewer must have sufficient graphicacy and other literacy-based skills draw upon to know what is missing, deduce what should be there, and hold that implied, invisible information in their working-memory as they continue to interact with the HDV.

For example, health data visualizations intended to be used by a wider audience, such as a BMI chart, often have missing personal information that the viewer is supposed to mentally fill-in-the-blank with their personal data to effectively read it. In this case, viewer-specific health information, like someone’s weight and height, are used to guide the viewer to their BMI number in the table.

One kind of fill-in-the-blank barrier is abbreviating content (i.e., Monday = Mon = M). Abbreviating content often requires the viewer to have prior knowledge of cultural or design norms to accurately interpret and effectively leverage the information presented. Using abbreviations to simplify the HDV’s design is a common design choice in data visualizations, especially those intended to be seen on small screens (e.g., a smartphone or wearable device). Single letter abbreviations caused fill-in-the-blank barriers for some people with Down Syndrome. In the “Daily Steps” bar graph, several participants struggled with the “k” abbreviation for kilometers. At first, Cameron thought the K meant one thousand; while Jesse thought it was a kilogram. Similarly, Jordan and Shiloh both struggled to recognize the days of the week when they were abbreviated to a single letter (i.e., SSMTWRF). In these examples, the abbreviation forced viewers to use additional context clues or have pre-existing knowledge to fill-in-the-blank.

Missing information is another kind of fill-in-the-blank barrier. When the missing information is related to a health topic that the individual is less familiar with, health data viewers with Down Syndrome may struggle to recognize that effective interaction with visualizations first requires information that may not be visible. When a fill-in-the-blank barrier occurs due to missing information, the viewer must first recognize the absence of crucial information before they can connect the components of the graph together. For example, someone viewing the “Body Fat for Men” table needs to know three key pieces of personal health information: their biological sex, age in years, and—crucially—their body fat percentage. If a viewer does not realize that additional information is needed to actually read a HDV, they will struggle to use it during later stage activities, like using the data as a health management decision-making tool.

Gaps in prior knowledge can also lead to fill-in-the-blank barriers. None of the participants recognized that they needed to know their specific body fat percentage to read the chart and find out where they fell across the four categories. This may have occurred because the participants were less familiar with “body fat” beyond an association with their weight or anatomy. Their lack of specific medical jargon appeared to make the numbers and topic more abstract and ambiguous. When this occurred, participants attempted to fill-in-the-blank by drawing upon their personal health literacy and knowledge. For example, Cameron interpreted these numbers as being related to “the scale ... the different numbers of wei–of weight.” Darcy viewed the values as “the number of pounds of your body fat” rather than the percentage. In these examples, personal health knowledge influenced and revised their mental models about the data being presented, which caused conceptual mismatches in later data visualization reading phases.

Some fill-in-the-blank barriers also occurred due to known numeracy issues (described in 5.1.2) as people with Down Syndrome can struggle visualizing a “mental number line” [39]. There were two common numeracy-based design causes for fill-in-the-blank barriers in this study’s HDVs: number ranges and intervals. The use of number ranges and numerical increments on the X and Y-axes caused confusion for multiple participants due to the invisible number lines implied by these commonly used graph design choices. Both number ranges and intervals require the HDV viewer to be able to mentally imagine an invisible number line. For example, “18–25” is a number range that abbreviates the full number line of 18, 19, 20, 21, 22, 23, 24, 25. The numeracy convention of using an em-dash is intended to imply the invisible numbers in between 18 and 25. In the Body Fat for Men table, both Jordan (16) and Cameron (17) recognized that their ages were not visible. The age brackets began at the 18–20 range in the y-axis. Because this data visualization procedure was not explicitly stated, some readers may likewise struggle to find where they would fall within a number range. Secondly, number intervals similarly employ abbreviations to imply the invisible number line in a graph’s x- and y- axes. The number interval on the axis of 0, 5, 10, 15 indicate that the visual spaces in between intervals (i.e., the 0, 1, 2, 3, 4, and 5 are implied as existing between the first and second line).

Invisible number lines may be particularly frustrating when first reading an HDV because of the interaction between numeracy skills and working memory. Both numeracy and working memory are integral to a person’s ability to: (1) mentally imagine the number line being described, (2) determine if their age is within the number line or not, (3) perform these two steps until a match occurs, (4) mentally associate their personal invisible age to the appropriate design element, and (5) continue to hold that now invisible numerical information-design element association in their working memory to draw upon as they continue to interact and reason with the HDV. As the numeracy skills of people with Down Syndrome may lag an average two years behind their print literacy skills and they may also experience attention or working memory issues in their daily lives [10, 14], these kind of health data visualization numeracy-based fill-in-the-blank barriers can be especially challenging to overcome.

When some participants encountered unknown content (i.e., unfamiliar words or were uncertain how to progress in their task), several reacted by ignoring the unfamiliar content entirely. One example of this was when participants did not know the word “macronutrient.” Rather than using their preferred information-seeking method or asking its meaning, multiple participants ignored this keyword altogether. The additional information-seeking task appeared to increase their mental effort whilst they were already progressing through their identification stage tasks. Unfortunately, this avoidant behavior led to later problems as text or numbers that were ignored often remained unconnected with other information in the second and third stages.

Most participants encountered encoding barriers when they were making connections between the various HDV design elements and other relationships between the data. Appropriately layering encoding is critical to the successful interpretation of visual metaphors. The two most successful HDVs were the stacked bar graph (95%) and the scatterplot (100%). The icons in these visualizations appeared to be the most easily connected because of the redundant encoding between the graph elements that resulted in the bundled channels which better supported people with Down Syndrome’s graph reading abilities.

The scatterplot’s use of pictures of familiar foods and drinks with text labels next to the image made it easier to understand the abstract data through concrete examples. Furthermore, the data itself reinforced participant’s understanding of and familiarity with nutrition. This combination supported their ability to more easily categorize, group, and generalize about the data more so than any other HDV. The stacked bar chart similarly used familiar icons with labels in close proximity in a prominent location (i.e., directly underneath the large title). Even though the graph used layered color-encoding (i.e., status and categorical), the color icons were more clearly connected when they were superimposed atop the larger, colored rectangle (e.g., the green lettuce icon in the center of the green rectangle of a stacked bar).

Transposing content can also interfere with accurate encoding. Half of the participants verbalized a preference against vertical text, which was unnecessarily difficult to read as several had to physically turn their head sideways. Both line graphs required participants to read dates and times where the text was vertically transposed. This was made more complex because both graphs had repeating values that resulted in blocks of text. For example, in the first line graph the x-axis was a date separated in weekly increments. The dual y-axis line graph the x-axis recorded timestamps of activity intensity at 30-second intervals. The transposed text at intervals appeared to increase cognitive load because participants had to compare values in the graph area against the y- and then against the x-axis which required fine-detail level comparisons to notice the differences. Several participants instead opted to respond more generally by saying the month, which was the easiest to read while also scanning and comparing the information.

Transposing can also break visual metaphors when using icons or images. The vertical orientation of the bars in the second graph triggered an encoding error of the step metaphor for some participants. “I see that the steps from up here, you’re walking down the stairs and then from Monday, you’re going up the stairs, and Sunday you’re coming back down” Darcy. In this example, we can see how design choices mimicked a real-life activity (i.e., walking up and down stairs). The different heights of the tops of individual bars were reinforced by both the title text steps and the shoe icons. As the word steps is also a synonym for stairs, we can see how orientation of data can affect the understanding of visual metaphors. Because of this, a participant’s mental model of a HDV can shift from total number of steps taken across a horizontal distance to the number of vertical steps upstairs. In this way, we can see how design choices that force viewers with Down Syndrome to transpose data can affect their encoding of visual metaphors.

While we found that icons and images can support participant understanding, the selection of appropriate design elements that successfully reinforce the visual metaphor is critical. One reason the stacked bar chart was the highest performing HDV is perhaps because it used multiple levels of redundant encoding to more tightly bundle graph elements during interpretation. Its success suggests that unused visual properties should be used to redundantly encode the main data dimensions because bundling channels of the same information has been found to support faster, more accurate, and efficient meaning acquisition ([126], pg. 179). Bundling data visualization channels with appropriate graph type and information orientation can make or break visual metaphors. These findings suggest that designing layered encoding in HDVs must be done with careful consideration. As such, future HDV design features should clearly and explicitly call attention to these design choices and define the intended meaning between elements to ensure the accurate understanding of the visual metaphor.

Participants struggled the most during the final stage because of unsupported inference-making. While we saw that people with Down Syndrome are visually strong, synthesizing meaning from connections made about observed relationships both within and outside of the HDV can be difficult. For example, Skyler was confused when they encountered a bar with a shoe icon cut in half at the top in the Daily Steps HDV. The shoe icon was cut off to indicate a partial value. However, there was no visual indication of what a partial values means. In other words, the appropriate way to interpret the half shoe was implied, rather than made explicit. This impacted Skyler’s ability to accurately connect the icon with the intended meaning. “What does the shoes mean, though? One of the shoes was cut off, but I could still see it.” When participants were uncertain how everything was connected, they were unable to deduce the underlying takeaway of the data being visualized. Without interpretive supports integrated into the HDV, people like Skyler may encounter similar conceptual mismatches that can affect their ability to fully engage with their health information.

Feelings of uncertainty caused by barriers or misunderstandings in the two previous stages appeared to diminish participant self-efficacy in their graphicacy abilities particularly in this third stage. When inference-making was not supported, participants often relied upon existing knowledge and skills that they felt more confident with employing. Instead, several participants drew upon what they already knew about health, nutrition, and exercise science when they felt uncertain what something meant or what information should be connected together. As a result, participant confidence in more familiar health literacy skills versus those where they had poor self-efficacy (i.e., graphicacy skills) caused several people to inaccurately update their mental model and understanding of the HDV.

Although knowledge is constantly informing people’s interpretation and the subsequent inferences that they make, pre-existing knowledge related to the HDV topic plays a more significant role when people attempt to read beyond a visualization. Drawing upon outside sources of information to synthesize and contextualize the graph’s data within the viewer’s understanding is at the heart of the third graph engagement stage. Several participants tried to retrofit their current mental models of the HDVs to match their pre-existing health knowledge regardless of the goodness of the fit or not when they did not know how to make inferences. Including information present to support inference-making in the HDV is not enough. The design of future HDVs need to explicitly support the inference-making process of people with Down Syndrome. This could be done by features designed to reduce the procedural uncertainty and ambiguity about the relationships between HDV elements (e.g., a guided walk-through). Inference-making supports like this may also help viewers with Down Syndrome to connect information with external knowledge to leverage greater insights into the data.

Another reason why inference-making may be difficult during this stage is that health data visualizations in particular are intended to be used as part of an informed decision-making process that is driven by questions. It is through question generation that graphicacy tasks are contextualized, HDV reading goals are created, and appropriate outside knowledge schemas can be drawn upon as viewers make inferences. When questions are not an integrated part of the HDV reading process, knowing what types of inferences are relevant to their task goals becomes more unclear. While we did investigate the kinds of questions participants ask when reading graphs, our findings suggest that future HDV features could better support question generation. For example, providing a bank of relevant questions that would guide viewers throughout the reading process would tie interpretation and inferences to that question’s goal (e.g., Am I getting enough protein this week? Am I walking enough each day?). Furthermore, question generation could allow for more of a narrative interpretation about the data and provide relevant takeaways as the viewer is guided through each stage. As past work has suggested that storytelling can be beneficial to data visualization, question-driven storytelling may provide greater opportunities for insight with HDV viewers with Down Syndrome in the final stage of graph reading. Integrating step-by-step graphicacy educational supports, like guided walk-throughs, could benefit people who may not know or be able to recall (i.e., inadequate education or lapsed graphicacy skills) what next steps to take when interpreting visualized information.

One way to address the myriad task-based barriers discussed in this section is to provide procedural supports to viewers with Down Syndrome. Future health systems could include a feature that visually models graph reading procedures in a sequential manner. The step-by-step walk-throughs could provide procedure supports and structure to the graph reading process, which the variety of saliency between participants indicate they could benefit from. Set by the pace of the viewer, this gradual approach to information acquisition may support the mental model construction process by offloading the cognitive effort required during unstructured graph reading. Furthermore, asset-driven guided walk-throughs could leverage their strengths as learners and provide support for any population weaknesses (see 2.1).

For example, highlighting content step-by-step and explaining what it means plays to their visuo-spatial strengths and preferences towards behavioral modeling, guides their attention to aid focus, supports encoding of meaning to ensure more accurate understanding. It also reduces information into smaller chunks that slowly build upon each other support working memory differences. The explicit meaning mapping between HDV elements and design choices, including providing additional health information, like definitions or health literacy explanations, can reduce task-shifting from the HDV to information-seeking activities outside of the HDV. Guided walk-through features may provide asset-driven supports to the short and working memory of individuals with Down Syndrome, as past work has noted population deficits in these [8]. With more short-term and working-memory freed up, viewers with Down Syndrome may be able to more deeply engage with visualized health information. Although interactive guided walk-throughs may be supported by some of this study’s findings as well as various strengths and weaknesses of learners with Down Syndrome (see 2.1), any proposed HDV feature ultimately requires a future participatory co-design study with and critical evaluation by people with Down Syndrome, which we will explore in future work.

As most mathematics literature for this populations orients around developing critical real-world numeracy skills, other skills that are becoming increasingly necessary in society, like reading data visualizations, have been overlooked. This study takes preliminary steps in better understanding the health data visualization literacy skills of individuals with Down Syndrome. While this study may have implications for educators, these are suggested starting points that should be taken with a grain of salt given the qualitative nature of this study’s format and small sample size. We broadly describe some approaches educators could take when looking to develop the graphicacy skills of students with Down Syndrome.

During lessons, one way to engage students with Down Syndrome or other IDDs is to play to their strengths as learners. As people with Down Syndrome have strong visual awareness, benefit from behavior modeling, are great kinesaethetic learners, and have an active limbic system, graphicacy skill development could be improved by making data activities be embodied experiences, more socially engaging, and emotionally evocative to facilitate the transition of graphicacy skills from the classroom into long-term memory. It may be valuable to use some of the identified barriers (summarized in Table 7) during lesson planning, so students with Down Syndrome can co-develop personalized graphicacy strategies when they encounter any of data visualization barriers described.

The questions in Tables 14 and 15 in the appendix may be useful starting points when initially evaluating the graphicacy skills of students with Down Syndrome. For example, asking them to describe data visualizations can provide insights into what they are first drawn to, may indicate what underlying literacy skills they are stronger or weaker in, what they think is important and relevant within the data visualization, and how they orient themselves within the various elements. Ideally, with graph-type specific training, this method of describing graphs may become more effective and standardized, indicating the transition of graphicacy skills from classroom activities into long-term memory. Evaluation of graphicacy skill development could also include asking specific questions looking for specific values, like identifying extremes. Asking students what kinds of differences they see can indicate their underlying pattern-finding strengths and help elucidate their comparative abilities to notice when data is similar or not. This may be helpful if students need help mapping abstract mathematics “type” language to visual elements (e.g., mapping descriptive superlatives like most to the largest values in the data visualization). Using specific, concrete token-based language to describe abstract mathematics language and concepts may also support this mapping according to some students with Down Syndrome’s type-token distinction preference.