Assessments can be used for many reasons, but something they almost all have in common is reporting ability, skill, or proficiency in some sort of metric, like a score. Commonly used data displays for reporting scores, such as line and bar charts, are recommended because people tend to be most familiar with these types of displays (Franconeri et. al, 2021), but their overutilization can actually lead to misrepresentation of the data (Holder, 2022) and false understandings when viewed by policymakers, educators, parents, and the general public. Figure 1 shows an example of how complex data can be oversimplified, leading to an inaccurate interpretation. While the reality of social data is frequently a series of overlapping distributions of different groups, when these data are expressed in an oversimplified but easy to grasp graph, like a bar chart, the truth of overlapping distributions is lost, and these data are perceived to be distinct narrowly defined groups with little overlap – a huge distortion of reality.

Figure 1

Example of how oversimplified data can lead to misinterpretation of the data

Copied from Holder (2022). While the reality of social data is frequently a series of overlapping distributions of different groups, when these data are expressed in an oversimplified but easy to grasp graph, like a bar chart, the truth of overlapping distributions is lost, and these data are perceived to be distinct narrowly defined groups with little overlap – a huge distortion of reality.

Just as humans are complex, any assessment meant to measure and represent human knowledge, either looking at one individual or a population, should show the breadth and depth of that complicated truth. Common score results displays typically only present a mean of a subgroup populations; however, simplistically viewing scores this way tells you very little. What is the range of scores? Are they evenly distributed or sporadically dispersed? How much overlap between the groups is there? How many points are in each group (which tells you something about measurement error)?

Figure 2

A current NAEP achievement gap display

Copied from NAEP (n.d.). Presenting achievement gaps between racial/ethnic subgroups like this is problematic, as it can lead to attribution bias, which is when a person attributes characteristics (typically negative) to an entire group of people, resulting in biased interpretations and stereotyping (Holder & Xiong, 2022).

For example, Figure 2 shows a current NAEP achievement gap display between public and nonpublic school students. It tells you the magnitude of disparity between the two populations and that it’s consistent over time. As implied by the misperception of data shown in Figure 1, presenting achievement gaps between racial/ethnic subgroups like this is problematic, as it can lead to attribution bias, which is when a person attributes characteristics (typically negative) to an entire group of people, resulting in biased interpretations and stereotyping (Holder & Xiong, 2022). In an example of this, Holder (2022) asked people to look at simple, frequently used data displays like bar and line charts for social data to see how they interpreted what they saw, such as the display in Figure 3. He found that people more frequently used a deficit frame to explain differences between groups, that is, attributing deficits to people, not structural inequalities. For example, when presented with the bar chart below, people explained the deficit in the pay of Group B due to lack of hard work or friendliness, rather than a structural cause, such as Group B working in a less busy or lower living-wage location.

Figure 3

Common display example that results in deficit framing

Copied from Holder (2022). People more frequently used a deficit frame to explain differences between groups, that is, attributing deficits to people, not structural inequalities.

To take this a step further, there seem to be a lot of researchers who have criticized achievement gap reporting in general. For example, Ladson-Billings (2006) states, “when we see the ‘achievement gap’, what we are really seeing is generations of inequalities that have compounded. Rather than gap, this is better understood as a debt that should be owed to those whom these inequalities have affected.” In this sense, achievement gap displays are not showing score differences, but rather, compounding inequality between races and ethnic groups that, as Figure 2 shows, has not changed much in the last 30 years. Policymakers, educators, parents – no one has been able to move the needle on eliminating structural inequalities in education. Why? Talking directly with teachers and parents is likely a good start, but a lot can also be done to improve score result displays to make it easier to investigate data, more deeply (from national, to state, to city, to school, and even classroom levels) and more broadly (over time, comparing specific indicators across states). To this point, I have a personal note: While I started on this journey being primarily interested in creating better visualizations than line and bar charts, I’ve quickly come to realize how poorly used large-scale assessment data are utilized, and when they are, the way in which results are portrayed can actively do more harm than good.

In this paper, I will start by discussing some of the harmful effects of standard large-scale assessment reporting (particularly achievement gap reporting) and explore why it is important not to oversimplify complex data. From there, I will share critical perspectives on score reporting best practices that were obtained from interviews with National Assessment for Education Progress (NAEP) data users (Kerzabi et al., 2024; 2025). Finally, I will explore some examples of quality score data reporting for different stakeholder groups and offer actionable suggestions to better support users with low data literacy.

There are a lot of best practices for visualizing data out there (e.g., Ancker et al., 2023; Franconeri et al., 2021; Holder & Xiong, 2022; Lane & Sandor, 2009; Padilla et al., 2018; Schwabish & Feng, 2021), and plenty of literature (experiments, interviews) about how various groups understand data (e.g., Fansher et al., 2022; Houston & Henig, 2023; Miller & Watkins, 2010; Zapata-Rivera et al., 2013), and there are plenty of sites that offer insights into the types of visualizations that would work best for different data (e.g., Data Viz Project, n.d., shown in the video below; Data Visualization Society, n.d.). Across the literature, four major recommendations surface when preparing data for visualization: (1) keep it simple, (2) keep it clean, (3) know your audience, and (4) manage the message.

Keep it simple

1. Use formats that are familiar to users (Franconeri et. al, 2021); use histograms or scatterplots over statistical summaries and jitter plots over bar charts to show variability (Holder, 2022; Holder & Xiong; Schwabish & Feng, 2021).
2. It’s also important to respect common associations, such as up for more and down for less, opaque or darker for intensity/more (on a light background), right to left for passage of time (Franconeri et. al, 2021).
3. Avoid taxing a user’s working memory (Franconeri et. al, 2021; Padilla et. al, 2018; Khasnabish, S., et al., 2020; Lane & Sandor, 2009); avoid statistical jargon and choose language carefully, particularly when communicating technical concepts (Goodman & Hambleton, 2004).
4. Avoid the trap of false simplicity – means mislead, show variability among subgroups (Holder, 2022; Holder, E. & Xiong, C., 2022).

Keep it clean

1. Avoid unnecessary embellishments (Franconeri et. al, 2021).
2. Convert legends into direct labels and integrate relevant text into visuals as direct annotations (Franconeri et. al, 2021).
3. Avoid error bars for general audiences, which can be misinterpreted as data ranges; rather, show examples of discrete outcomes (simultaneously or over time; Franconeri et. al, 2021).
4. Exclude or reduce gridlines (Khasnabish, S., et al., 2020) and avoid shaded backgrounds (Lane & Sandor, 2009).
5. Round to whole numbers where logical, otherwise not more than two decimal places (Wainer, 1992).

Know your audience

1. Rely on absolute values instead of relative rates when your audience has issues with numeracy (low numeracy; Franconeri et. al, 2021)
2. Familiar formats may vary by the intended audience (Franconeri et. al, 2021)
3. Combine graphical and numerical displays to support audiences with varying or unknown numeracy levels (Lane & Sandor, 2009; Khasnabish, S., et al., 2020).
4. Convey probabilities with frequencies instead of percentages and use well-constructed icon arrays with the same denominator (Franconeri et. al, 2021).
5. It may help to consider, “what do I want the recipient to think, feel, or do, after receiving this information?” (Ankler et. al, 2023; e.g., reporting trends over time may be for the purpose of recall, displaying the data so it can be remembered, while reporting differences between subgroups may also have an influence on one’s perception of those subgroups).
6. Avoid pairing red and green, pair red and blue instead, and select different scatter-point shapes for further accessibility (Franconeri et. al, 2021; Padilla et. al, 2018).

Manage the message

1. Know your desired outcome before designing a visualization (Ancker et al, 2023).
2. Use visual grouping/clustering cues to control which sets of comparisons (subsets of values) a viewer should make and use annotation and highlighting to narrow that set to the single most important comparison that supports your message (such as in a presentation of results; Franconeri et. al, 2021; Khasnabish, S., et al., 2020; Lane & Sandor, 2009).
3. With subgroups, use jitter plots over bar plots with median line and show prediction interval rather than confidence interval to reduce stereotyping (Holder, 2022; Holder, E. & Xiong, C., 2022; Lane & Sandor, 2009).
4. Consider the color and hierarchy implications of your legend (Schwabish & Feng, 2021)

Complex data can be clarified through language, data grouping, the ways in which variability, significance, and uncertainty are relayed, and even through how data legends are displayed. Figures 5 – 11 offer some examples of how to convey such information cleanly and clearly, while still retaining data complexity.

Figure 5a

Using shape, color, size, and position to convey meaning

Copied from Franconeri et al. (2021). Grouping cues like color, shape, size, and position can all control how data are conveyed.

Figure 5b

Impact of using grouping and color

Copied from Franconeri et al. (2021). By grouping and coloring by type of earnings rather than by year, the difference by year is emphasized.

Figure 6

An example of the importance of displaying variability in data

Copied from Schwabish and Feng (2021). In this graph, a large amount of variation would have been lost had only the mean (black circle) for the major groups been presented. For example, although American Indian and Alaskan Native is combined as one group, poverty among Aleut Alaskan natives is far below the mean for the group overall; similarly, those identifying as two or more races include a wide range of race/ethnicity combinations.

Figure 7

Example showing how different displays of uncertainty can change the message of the data

Copied from Franconeri et al. (2021). Original source and description embedded in the figure, e.g.: “The chart on the right is the same as the chart on the left, except that it has the y-axis range of the chart in the center… viewers are sensitive to superficial cues, such as how big a difference looks within the frame of the chart.”

Figure 8

Additional considerations for displaying data uncertainty

Copied from Franconeri et al. (2021). Original source and description embedded in the figure, e.g., on the left, “fuzziness, location, arrangement, and color value were rated as relatively more logical than other variables for representing uncertainty;” on the right, a color map “encodes value (via hue) and uncertainty (via lightness).”

Figure 9

Considerations for formatting legends to avoid racially biased implications

Copied from Schwabish and Feng (2021). Original sources embedded in the figure. Color palettes should be chosen to avoid color-associated implications and hierarchies. The panel on the left is a poor example, in which the color black is associated with associated not only with race but also with poverty. The panel on the right is gradient and hierarchy neutral without implicit associations.

Figure 10

Considerations for naming and explaining graphs and figures

Copied from Schwabish and Feng (2021). Original source embedded in image. The right frame title and subheading more accurately shares the implication of the data for general audiences.

From a fairness and equity standpoint, Figure 5 is particularly important, as it shows how easily data can be misrepresented in seemingly “neutral” frames. In left panel, the title and subheading appears clinical, invoking the fallacy that numbers are neutral (Castillo & Strunk, 2024). While likely not intentional, a lay person may not understand the implication of the data. Anyone working in a prison will know that a mental health diagnosis is generally considered a positive thing (Gómez-Figueroa & Camino-Proaño, 2024), in that it opens the door to necessary treatment and symptom management while incarcerated. With the data shows that a larger majority of white inmates are diagnosed with mental health disorders than people of color (POC), the implication is that white inmates have disproportionate access to mental health treatment, effectively, racism in jail. The right frame title and subheading more accurately shares the implication of the data for general audiences.

While this section provides general best practices alongside examples, even best practices can be impractical for data users. In this section, I share some criticisms directly from stakeholders that I helped to interview on this topic.

In this section, I share some criticisms directly from stakeholders that I helped to interview (see Kerzabi et al., 2025) in January and February of 2024. My team and I conducted interviews with NAEP state coordinators (people that liaise between the US Department of Education and their State government), as well as a few policymakers (elected state officials) and their staffers. When presented with some of the recommendations noted in the previous section, here were some of the comments we received from interviewees.

Recommendation #1: Show a range of data when reporting (AERA, APA, NCME, 2014); don’t presume white students or males as the focus group (Schwabish & Feng, 2021)

• Benefits: With this type of display, you still see that there are score differences, but the range represents the entire distribution of the groups scores in a more realistic and honest way. (Note that the APA testing standards also call for contextual statements to be included whenever possible as well, but it seems like such statements are either not included or not included in any meaningful way.)
• Example: In the left panel, you see the traditional achievement gap display (comparison black vs. white focal group); in the right panel, you see the range of students in the subgroup group compared to all students.

Figure 11

Gap vs. range display

Copied from Kerzabi et al. (2024).

• Criticisms:

"It’s difficult to make the public understand graphs with distributions. Same with percentiles." (from Kerzabi et al., 2025, p. 13)

"People want to know the bad story and try to fix it." (from Kerzabi et al., 2024, p. 16)

Recommendation #2: Conduct within subgroup analyses to see what the best achievers within that group are doing (Hughes, 2023).

• Benefits: Find best practices for supporting low achievers within that same group; sometimes also called asset-based reporting or opportunity to learn.
• Criticisms:

“There’s a reason why low achievers are hindered. [We] can’t think that doing what we do with high achievers will bring everyone up, and people might find it condescending.” (from Kerzabi et al., 2024, p. 37)

“Digging into why that context exists is pretty complex and can take time… you may not know exactly what the responses meant or why a particular state/district/region responded that way.” (from Kerzabi et al., 2024, p. 26)

“Too often [opportunity to learn] is simplified as poverty, minority, or zip code. Sometimes it… ignores home life when it should, in fact, be much broader than just school opportunities.” (from Kerzabi et al., 2024, p. 39)

Recommendation #3: Keep data visualizations simple (Khasnabish et al., 2020; meta-analysis of data visualization in healthcare)

• Benefits: Simple data displays are quick to interpret and easy to remember (for the elevator speech) and generally point to specific actions that should be taken based on the data. It is also fairly well-established that data literacy varies among members of the general public, as well as among those in positions that affect policy change.
• Criticisms:

“It’s always better to tell a complex story that’s true than a simple story that’s false.” (from Holder, 2022)

“The thing people most want is more detailed data, they want more in depth data.” (from Kerzabi et al., 2025, p. 20)

“There’s [a lot of data] there, they just don't know how to use it or interpret it.” (from Kerzabi et al., 2025, p. 20)

To summarize, there’s a definite conflict between wanting a lot of information to be able to dig deeper into the data and more accurate displays, while also wanting clear and actionable information to be taken from displays. This is discussed in the final section.

Finally, I would like to share some ideas for different types of data displays that serve a useful purpose for the intended audience (general public, policy makers, teachers and parents). Building off of last week’s update, how can student data be presented holistically, while still being understood by general users. Figures 12 – 15 are a few displays for the general public on large-scale assessment data that are meant to provide additional context, score distributions, etc.

Figure 12

A display on the impact of COVID-19-associated job loss by race/ethnicity

Copied from Schwabish and Feng (2021). Original source embedded in the figure. This figure offers an example combining recommendations from Figures 8 and 9, selecting colors free of biased associations and transparency encoded for specific meaning; here, transparency represents statistical significance (rather than gradients of uncertainty).

The Programme for International Student Assessment (PISA) must frequently show population differences, and as such, usually does a good job of using dispersions when presenting multiple group visualizations. The example in Figure 13 shows mean performance of the different nations, and the dispersion of economic, social and cultural status (ESCS index) is dispersed.

Figure 13

Example of dispersion comparing different subgroups from PISA

Copied from Schleicher (2018)

Figure 14 shows a created display to illustrate the impact of SES on student outcomes. Notice how the choice of color keeps grouping variables aims to keep results stigma- and bias-free.

Figure 14

Score distribution example highlighting structural inequalities

Copied from Kerzabi et al., 2024. In this figure, color associations are free of bias (as in Figure 9) and the size or width of the bar carries meaning as to the size of that portion of the population (as in Figure 5a).

Figure 15 offers an alternative to achievement gap reporting; rather than pitting groups against each other, growth toward a fixed goal is displayed. Note that this display is particularly clean, without extraneous gridlines (Khasnabish, S., et al., 2020) or a shaded background (Lane & Sandor, 2009). As with other trend graphs, note how this display also maintains common associations, such as up and down for more or less and left to right for the passage of time (Franconeri et. al, 2021).

Figure 15

An achievement gap alternative for comparing subgroups

Copied from Blakely (2019). This graph offers an alternative to achievement gap reporting, comparing groups to a standard, rather than to each other. Further, this display is particularly clean, without extraneous gridlines (Khasnabish, S., et al., 2020) or a shaded background (Lane & Sandor, 2009).

So far, these displays have focused on equitable representations of group score data; however, another major concern when displaying data is making displays interpretable for users with low data literacy. The next few displays (Figures 16 – 17) may be useful for these general users, who may have low data literacy. These graphs avoid statistical jargon (Goodman & Hambleton, 2004) while putting the data into its appropriate context. Through simple images and icons with descriptive text, interpreting the information is less taxing for users who may not be able to accurately decipher even common graphs (Franconeri et. al, 2021; Padilla et. al, 2018; Khasnabish, S., et al., 2020; Lane & Sandor, 2009). Formats such as an a speedometer or a gas tank gauge (with an arrow rotating from left for slow/empty to right for fast/full), as well as a commonly used bar chart for cell phone service (with increasingly high bars filling as the phone gets better reception) are familiar and common to users, making them readily applicable with new information but with similar associations (Franconeri et. al, 2021).

Figure 16

Example of the California State Education dashboard for users with low data literacy

Copied from California (n.d.). Note the speedometer / gas tank gauge (with an arrow rotating from left for slow/empty to right for fast/full), as well as a commonly used bar chart for cell phone service (with increasingly high bars filling as the phone gets better reception) are familiar and common to users, making them readily applicable with new information but with similar associations (Franconeri et. al, 2021).

Figure 17

Example of the Delaware Report Card for users with low data literacy

Copied from Delaware (n.d.). Through simple images and icons with descriptive text, interpreting the information is less taxing for users who may not be able to accurately decipher even common graphs (Franconeri et. al, 2021; Padilla et. al, 2018; Khasnabish, S., et al., 2020; Lane & Sandor, 2009).

For users that really want to do a deep dive on data, such as parents interested in knowing specifically about their own child, or a teacher about their students (and whose pay may be tied to the academic performance of their class), or a principal or superintendent about their school or district, complex data can still be displayed in an understandable context using words, colors, and common associations. Some institutions, such as the California Department of Education, have created training videos to support novice users, as well as promotional videos to increase familiarity and accessibility. Videos 1 and 2 are both examples of instructional videos.

Video 1

Around the California Dashboard in 90 seconds

Source: California Department of Education (2019)

Video 2

Introductory video to the MAP Growth 101 YouTube playlist, with each video exploring a different and important area of understanding and interpreting student growth data.

Source: NWEAvideos (2024)

The Northwest Evaluation Association (NWEA) has done a nice job of expressing data effectively, as well as making their growth reports accessible to parents and educators. Their introductory overview is shown in Video 2, and the rest of this section looks at some of their reports. Figures 18 – 20 show the NWEA “Map” example reports for an individual student, a class (achievement and growth), and at the school or district level, respectively. Note how these combine graphical and numerical displays (Lane & Sandor, 2009; Khasnabish, S., et al., 2020) and show clear values rather than frequencies or rates (Franconeri et. al, 2021). Figure 18, in particular, offers clear and concise information about the student and seems mindful that the parent will be interpreting this information (Ankler et. al, 2023).

Figure 18

NWEA Family Report (for a student)

Copied from NWEA (n.d.).

Figure 19

NWEA Achievment Status and Growth Report (for a class)

Copied from NWEA (n.d.).

Figure 20

NWEA School Profile

Copied from NWEA (n.d.).

Many of these displays do not incorporate all best practices, but they seem to intentionally incorporate many. With these final displays shown in Figures 18 – 20, it should also be pointed out that these displays are also interactive, allowing the user to zoom in and zoom out on details to get more specifics of the score when interested. Interactive features can take data displays to the next level, and more explicitly support distinctly different users with varying degrees of data literacy.

In the next section, I will discuss and summarize the major findings of this paper.

As motivation for this paper, the introduction took an equity standpoint and looked at how complex data (such as group-score and large-scale assessment results) can be easily misrepresented and incorrectly understood when oversimplified and expressed using common data displays. Such data representations can paint a false narrative, which, while unintentional, can lead to attribution bias and harmful stereotyping by those looking at the data. The next section reviewed score reporting best-practices, including best practices among data equity researchers. These practices can be distilled as keep it simple, keep it clean, know your audience, and manage the message. Drawing on real conversations with NAEP data users, the third section shares specific critiques of these best-practices from those who work with large-scale, group-score data most frequently and for varying purposes. Finally, in the last section, equity concerns, best-practices in score reporting, and user critiques were synthesized to present some good examples of how data can be visualized with different users in mind.

It is my hope that this paper will support those working with achievement data as both a reference and guide to creating equitable, interpretable, and useful data visualizations. While this paper targeted primarily summative assessment data displays, formative assessment displays are close on the horizon as technology and AI are more and more incorporated into daily classroom practices at-large. For these displays, the four main takeaways still apply, and I might add that, in addition to managing the message, knowing the question the data display should answer will be highly important. Simple, at-a-glance performance gauges, such as those in Figures 16 and 17, as well as displays that allow for easy drill-down for teachers, should be prioritized when developing formative assessment dashboards.