中国365bet中文官网

Identify biased data

Background:

Data visualizations can significantly affect how people understand and interpret data. But data visualizations can be biased and exclusionary, perpetuating inequity and harmful stereotypes. Discrimination can arise from how people use and misuse data, including which data they focus on or ignore. It's important to remember that numbers are a proxy for real people, places, and things and to respect and support the people behind the data. This means considering the data's context and the potential harm or erasure that may result from how they are presented.

People often view data visualizations as , but they are always shaped by the social and cultural contexts in which they are created. The 2021 call to action emphasizes that it is the responsibility of authors, editors, and reviewers to ensure that their publications effectively address social inequities and power dynamics in their analyses and findings. In practical terms, this means acknowledging any limitations or concerns with the available data and interpreting and contextualizing them for a more nuanced and complete understanding of the subject.

Caution is needed for law enforcement statistics in particular, as they can both . In the book , authors Mariame Kaba and Andrea J. Ritchie explain that police often exaggerate crime data and statistics to construct a narrative of "out of control" crime and to justify expanding police powers and budgets. In a , writer Matthew Hutson says, 鈥淧erhaps, whenever someone offers up an especially compelling explanation for a rise or fall in crime, we should be wary. . . . In hearing an explanation for its rise or fall, we might ask: What kind of story is its bearer trying to tell?鈥�

Incorporating a diverse range of sources and engaging with the most affected people can ensure that data reporting and visualization are fair and inclusive and that communicators don鈥檛 disadvantage historically excluded communities. Additional resources, such as the Urban Institute鈥檚 and the National Network of Education Research-Practice Partnerships鈥� guide can provide further guidance on this topic.

Recommendation:

Examine the data for biases or gaps. Ask whether the issue being analyzed and visualized is and what factors contribute to inequities. Seek a diverse array of sources to provide a balance of experiences and perspectives, and try to incorporate the voices of those most affected by the data. The 2019 DataJournalism.com article by Kim Bui suggests questions that may help evaluate data, including 鈥淲ho is vulnerable in this story and how would they want to be counted?鈥� 鈥淲ho is undercounted or possibly missing entirely?鈥� 鈥淲ho did the counting?" And 鈥淲ho benefits if you forget the dots are people?鈥� To help people understand how conclusions were reached, be transparent about the methodology used, and provide text that explains the systemic causes of differences shown in a visualization.

Use caution when reporting on crime and public safety. Avoid relying solely on law enforcement statistics as sources, and be cautious when using aggregated data. for data from government or corporate officials so that your audience better understands the complexities of the data.

See also 鈥淚nvolve a diverse group of people in the creative process,鈥�.

Design with empathy

Background:

Commonly used forms of data visualization, such as bar charts, line charts, and pie charts, while informative, can abstract content by collapsing the people represented into one shape, making it difficult for audiences to connect with the data on a human level. In a , software architect Jacob Harris discusses the difficulty of maintaining this perspective when dealing with large datasets: 鈥淯nfortunately, many data journalism examples focus exclusively on the far distance and leave out the near view. . . . As tools have improved, it has become phenomenally easy to put a bunch of dots on a map or in a chart, yet the legwork of understanding the 鈥榥ear鈥� of that data remains just as time-consuming. Under deadline time pressure, it鈥檚 easier to just plot the map and call it a day. But we lose something in the process.鈥� One way to show both 鈥渘ear鈥� and 鈥渇ar鈥� data is to pair patterns with deeper dives on subgroups in the data. Data visualizations should be grounded in empathy and lived experiences to help audiences better relate to the people behind the data. The use of design strategies to help audiences relate to people in data has recently been termed anthropographics. In a , researchers Luiz Morais, Yvonne Jansen, Nazareno Andrade, and Pierre Dragicevic identify several factors that may affect the ability of a visualization to elicit empathy or compassion from the audience, including the number of people represented by one shape or graphical mark and how completely each item in a chart can be distinguished from another.

Recommendation:

Use "near and far" graphics, which allow people to understand data on both a macro- and microlevel. Avoid focusing exclusively on the far distance and leaving out the near view. Consider whether certain chart types better lend themselves to connecting with the human element of the data or whether focusing on a smaller range of data might help people understand the personal stories behind the numbers. The use of icons to represent data points instead of abstract shapes such as circles can also help remind audiences that the data are not just numbers or statistics. Other techniques that can be used to help humanize data in tandem with visualizations include photography, profiles, illustrations, pull quotes, and oral histories.听

Examples:

Disaggregate when possible

Background:

Aggregating data鈥攃ombining multiple groups鈥攃an oversimplify data or , leading to potentially biased or misleading analyses and a for the groups that need it most. Examples of aggregation include , combining all American Indian and Alaska Native people, and grouping genders into only binary male and female groups. An example in which aggregation leads to inaccurate conclusions is when all Asian populations are included in a single group. In US science, technology, engineering, and mathematics (STEM) programs, this aggregation can mean that Southeast Asian people鈥攊ncluding Hmong, Vietnamese, Filipino, Laotian, and Cambodian people, among others鈥� even though STEM fields have historically excluded these subgroups.

Aggregation is particularly problematic when groups are lumped into a category called 鈥淥ther,鈥� as that label is ostracizing. Alternatives that the Urban Institute's recommends for racial and ethnic categories include 鈥渁nother race,鈥� 鈥渁dditional groups,鈥� 鈥渁ll other self-descriptions,鈥� 鈥減eople identifying as other or multiple races,鈥� 鈥渋dentity not listed,鈥� and 鈥渋dentity not listed in the survey.鈥�

In contrast, data disaggregation鈥攕eparating data鈥攃an highlight the specific needs of distinct groups. For example, disaggregated data in San Francisco County shows that , showing the need for targeted support. The National Congress of American Indians, the Asian and Pacific Islander American Health Forum, the , and others have called for data disaggregation as a way to promote equity. 听

One challenge to data disaggregation is small sample sizes. But instead of omitting these groups altogether, the in a 2022 article that communicators can present the data with explanations about lower reliability of data or show the group name with empty data.

One of the ways to support data disaggregation is to ensure robust and inclusive data collection practices. For example, people can allow write-in options for people to self-describe their identities and do outreach to ensure sample sizes will be large enough for statistical analysis. These practices should be conducted in collaboration with members of the community whose data are being collected. Collecting more data and for a wider range of groups can help research reflect the lives of all people.

Recommendation:

Disaggregate data when groups experience dissimilar effects, they don鈥檛 share a history, or members of the community say analyzing these populations together is unreasonable (). For example, when discussing data on gender groups, . Avoid combining data on distinct groups, such as data on Pacific Islander people and Asian American people. When you lack information for specific groups, acknowledge the limitations of the data. Also . If the use of a category comprising multiple identities is necessary, define the groups included, and use language that is respectful and descriptive rather than simply naming it 鈥淥ther.鈥澨�

Collect data in a way that can support disaggregation. And ensure that people can self-identify in a way that reflects their reality. Whenever possible, when collecting, analyzing, and reporting data to ensure the data accurately reflect the experiences and needs of the people surveyed and balance privacy concerns with the need for representation. See also 鈥淗andling small data samples.鈥�

Example:

Handling small data samples

Background:

Sometimes communicators omit data from a visualization because the sample sizes are very small. The National Congress of American Indians says American Indian and Alaska Native people may be described as the because many data visualizations use an asterisk to note missing or insufficient data on this group. But omitting groups from data is a form of erasure. Abigail Echo-Hawk, former chief research officer of the Seattle Indian Health Board, explains in a by writer Manola Secaira, 鈥淥ne of the ways that there is a continuing genocide against American Indians/Alaska Natives is through data. When we are invisible in the data, we no longer exist.鈥� Another example of erasure in data is when by gender. 听

When the sample size for a group of people is small, it is better to acknowledge missing or limited information rather than omitting the category altogether. Heather Krause explains in a , 鈥淭here is a world of difference between saying 鈥榳e鈥檙e less certain鈥� and 鈥榯hey don鈥檛 count.鈥� If you have a small sample size for a certain group, what you can say about them might be less 鈥榗ertain鈥� or 鈥榬eliable鈥� than you want, but that doesn鈥檛 mean we should discount their data with terms like 鈥榥ot statistically significant.鈥� 鈥� Noting missing or limited information acknowledges the absence of those communities and may encourage future inclusion. Using arbitrary rules to declare data "not statistically significant" , disregards useful data, and shifts the responsibility for creating meaningful results onto marginalized populations.

It is difficult to determine an exact threshold for when a group is too small to be broken out separately. 鈥淕roup size will always, however, be a practical consideration, which should be weighed against the severity of a subgroup鈥檚 social and health disadvantage and how different it is from the larger group, based on available information,鈥� says the report , by the University of California, San Francisco, and Robert Wood Johnson Foundation.

Showing small groups may also raise . 鈥淧articularly when subgroups are small (e.g., one American Indian Tribal Nation), concerns mount about being able to distinguish identities of individuals included in data collection. . . . The smaller a group, the easier it is to identify an individual, even without having names and addresses,鈥� Health Equity: Everyone Counts says. The Equitable Data Working Group also notes that .

Small sample sizes can be challenging to work with, but there are ways to maximize the data obtained from them. In a , researchers Carlotta Ching Ting Fok, David Henry, and James Allen discuss ways to optimize the data from small samples, including using specific research designs and statistical analysis techniques.

To prevent small sample sizes, data collectors must prioritize proactive outreach to groups that they have historically excluded. This work involves engaging with communities, organizations, and individuals to get more comprehensive and representative data.

Recommendation:

If the sample size for a particular group is very small, avoid omitting the data entirely or noting 鈥渘ot statistically significant鈥� without additional context. Instead, as long as privacy can be maintained, present the data while acknowledging the limitations鈥攆or example, , and provide measures like confidence intervals and standard deviations. Also, explore data collection and statistical techniques that will give analyses of small groups more reliability.

If presenting data on small groups would compromise people's privacy, an alternative approach is to present all category names in the same chart. Instead of displaying the data for too-small groups, display a message explaining why data cannot be shown. Communicators can also , such as text descriptions, if quantitative data are too unreliable and privacy is an issue. See also "Disaggregate when possible."

To prevent small sample sizes, data collectors should prioritize proactive outreach efforts, including building partnerships with community organizations and experts, using multiple recruitment channels, tailoring collection methods to engage underrepresented communities, and fostering trust and rapport with potential participants.

听

Examples:

Choosing color in data visualizations

Background:

鈥淚n data visualization, color sets the tone and enforces a message for the underlying visual display,鈥� contributor Eva Murray writes in a . Poor color choice can distract from content, confuse the meaning, and lead to data manipulation, intentionally or not. One way color can influence people鈥檚 interpretation of data is by its associations鈥攕uch as to certain political parties, brands, cultural traditions, and emotions.

According to a , English-speaking participants commonly associate the color red with concepts related to aggression, attraction, or heightened attention (e.g., stop, danger), while darker colors often have negative connotations. Consequently, using these colors to represent people can be problematic, as they may imply danger or negativity.听

Another consideration is avoiding colors that perpetuate stereotypes. For example, the Urban Institute鈥檚 recommends against using pink for women and blue for men鈥攁 combination that could 鈥攁s well as matching colors to skin tones (like using black to represent Black people).

Beyond the individual colors chosen, communicators should understand how the palette of colors may affect data interpretation. Using shades of a single color, for example, implies degrees along a spectrum, so that approach is better for data on a scale, such as temperature. Shades are less appropriate for categories such as races and ethnicities, . Understanding how combinations of colors can reinforce ideas about hierarchy and differences among social groups can help communicators show data with empathy.

Finally, all color choices should be accessible. For example, adjacent colors that have meaning should have adequate contrast against one another (see also 鈥淐olor contrast鈥�).

Recommendation:

Avoid choosing colors at random without considering their meaning or cultural associations, especially when the color's job is simply to provide distinction against adjacent content. Recognize when certain colors may perpetuate stereotypes. For example, when representing gender data, consider using a , as pink and blue do.

Use for the type of data being visualized, and prevent introducing distortions in the data. For more information on choosing colors for specific data visualizations, see and , which go into more detail on the most effective color palettes to represent different types of data. See also 鈥淐olor to convey meaning鈥� and 鈥淐olor contrast.鈥�

Ordering groups in data visualizations

Background:

The order of groups in a chart, table, or other visualization can imply a hierarchy or suggest that certain groups are the standard for comparison, which can lead to bias and . In the American Psychological Association notes that groups on the left of a graph or the top of a table may be perceived as more socially dominant or suggest that these groups are the standard for comparison. : powerful individuals or groups are placed above those with less power. In a , participants were asked to quickly judge pairs of powerful and powerless groups. Participants were faster at identifying the more powerful group when it was placed at the top. A separate found that people mentally represent a leader's power using vertical positions in space. The study revealed that even small differences in spatial elements, such as the leader's office floor level or the height at which their picture is displayed on a wall, can affect how power is perceived. Research has even suggested that . Considering the order and positioning in data visualizations and how they shape perceptions of power can help mitigate biases.

Recommendation:

Carefully think about the order in which groups are presented and how it might imply a hierarchy. Consider starting with the particular group the study is focused on or sorting the groups alphabetically. Interactivity can also be used to allow users to customize the order of presentation.听

Example:

Accessibility considerations for data visualizations

Background:

, and those that do can be confusing to learn. to people with motor disabilities, especially if they require zooming and navigating a dataset, clicking on data points or hot spots, or interacting with chart menus. Additionally, while it may be relatively easy to translate data describing frequencies of different categories or values from bar charts, pie charts, and line charts, .

Recommendation:

Ensure that interactive visualizations are , meaning users are able to navigate and interact with the data points, chart menu, and other chart controls using the keyboard only. Provide a text summary or data export of the visualization, making sure to describe trends or patterns in the visualization. that can be reasonably inferred from the visualization rather than presenting the raw and unaggregated dataset.

To make it easier to visually distinguish between stacked bars and points, it can be helpful to add a border to reduce any unintentional visual effects that may occur at intersections. In addition, label parts of the graphic directly when possible. If you use a key, order things in the key in the same way that they are ordered in the graphic itself. These strategies can help people associate visual elements with values. See also the 鈥淎ccessibility" section.

Resources on inclusive data visualizations

• Cesal, Amy. Nightingale, July 23, 2020.
  
• Dhawka, Priya, Helen Ai He, and Wesley Willett. CHI '23: Proc. 2023 CHI Conf. Hum. Factors Comput. Syst. article 807 (April 2023): 1鈥�14. https://doi.org/10.1145/3544548.3581086.
  
• Krause, Heather. We All Count, Dec. 4, 2020.
  
• Morais, Luiz, Yvonne Jansen, Nazareno Andrade, and Pierre Dragicevic. IEEE Trans. Visualization Comput. Graphics 28 (March 2022): 1661鈥�79. https://doi.ieeecomputersociety.org/10.1109/TVCG.2020.3023013.
  
• Muth, Lisa Charlotte. .
  
• Project Inkblot. The Creative Independent, Sept. 18, 2019.
  
• Reich, Julia. Stone Soup Creative, Dec. 29, 2020.
  
• Schwabish, Jonathan, and Alice Feng. . Washington, DC: Urban Institute, June 9, 2021.
  
• Urban Institute. .
  
• Urban Institute.
  
• Winter, Bodo, Sarah E. Duffy, and Jeannette Littlemore. Metaphor Symb. 35, no. 3 (2020): 188鈥�205. https://doi.org/10.1080/10926488.2020.1794319.
  
• Zhou, Li. Vox, May 5, 2021.