Principles of Effective Data Visualization (2024)

Box 1

Regarding terminology, the terms graph, plot, chart, image, figure, and data visual(ization) are often used interchangeably, although they may have different meanings in different instances. Graph, plot, and chart often refer to the display of data, data summaries, and models, while image suggests a picture. Figure is a general term but is commonly used to refer to visual elements, such as plots, in a scientific work. A visual, or data visualization, is a newer and ostensibly more inclusive term to describe everything from figures to infographics. Here, I adopt common terminology, such as bar plot, while also attempting to use the terms figure and data visualization for general reference.

Principle #1 Diagram First

The first principle is perhaps the least technical but very important: before you make a visual, prioritize the information you want to share, envision it, and design it. Although this seems obvious, the larger point here is to focus on the information and message first, before you engage with software that in some way starts to limit or bias your visual tools. In other words, don't necessarily think of the geometries (dots, lines) you will eventually use, but think about the core information that needs to be conveyed and what about that information is going to make your point(s). Is your visual objective to show a comparison? A ranking? A composition? This step can be done mentally, or with a pen and paper for maximum freedom of thought. In parallel to this approach, it can be a good idea to save figures you come across in scientific literature that you identify as particularly effective. These are not just inspiration and evidence of what is possible, but will help you develop an eye for detail and technical skills that can be applied to your own figures.

Principle #2 Use the Right Software

Effective visuals typically require good command of one or more software. In other words, it might be unrealistic to expect complex, technical, and effective figures if you are using a simple spreadsheet program or some other software that is not designed to make complex, technical, and effective figures. Recognize that you might need to learn a new software—or expand your knowledge of a software you already know. While highly effective and aesthetically pleasing figures can be made quickly and simply, this may still represent a challenge to some. However, figure making is a method like anything else, and in order to do it, new methodologies may need to be learned. You would not expect to improve a field or lab method without changing something or learning something new. Data visualization is the same, with the added benefit that most software is readily available, inexpensive, or free, and many come with large online help resources. This article does not promote any specific software, and readers are encouraged to reference other work14 for an overview of software resources.

Principle #3 Use an Effective Geometry and Show Data

Geometries are the shapes and features that are often synonymous with a type of figure; for example, the bar geometry creates a bar plot. While geometries might be the defining visual element of a figure, it can be tempting to jump directly from a dataset to pairing it with one of a small number of well-known geometries. Some of this thinking is likely to naturally happen. However, geometries are representations of the data in different forms, and often there may be more than one geometry to consider. Underlying all your decisions about geometries should be the data-ink ratio,7 which is the ratio of ink used on data compared with overall ink used in a figure. High data-ink ratios are the best, and you might be surprised to find how much non-data-ink you use and how much of that can be removed.

Most geometries fall into categories: amounts (or comparisons), compositions (or proportions), distributions, or relationships. Although seemingly straightforward, one geometry may work in more than one category, in addition to the fact that one dataset may be visualized with more than one geometry (sometimes even in the same figure). Excellent resources exist on detailed approaches to selecting your geometry,15 and this article only highlights some of the more common geometries and their applications.

Amounts or comparisons are often displayed with a bar plot (Figure1A), although numerous other options exist, including Cleveland dot plots and even heatmaps (Figure1F). Bar plots are among the most common geometry, along with lines,9 although bar plots are noted for their very low data density16 (i.e., low data-ink ratio). Geometries for amounts should only be used when the data do not have distributional information or uncertainty associated with them. A good use of a bar plot might be to show counts of something, while poor use of a bar plot might be to show group means. Numerous studies have discussed inappropriate uses of bar plots,9^,17 noting that “because the bars always start at zero, they can be misleading: for example, part of the range covered by the bar might have never been observed in the sample.”17 Despite the numerous reports on incorrect usage, bar plots remain one of the most common problems in data visualization.

Compositions or proportions may take a wide range of geometries. Although the traditional pie chart is one option, the pie geometry has fallen out of favor among some18 due to the inherent difficulties in making visual comparisons. Although there may be some applications for a pie chart, stacked or clustered bar plots (Figure1A), stacked density plots, mosaic plots, and treemaps offer alternatives.

Geometries for distributions are an often underused class of visuals that demonstrate high data density. The most common geometry for distributional information is the box plot19 (Figure1E), which shows five types of information in one object. Although more common in exploratory analyses than in final reports, the histogram (Figure1B) is another robust geometry that can reveal information about data. Violin plots and density plots (Figure1G) are other common distributional geometries, although many less-common options exist.

Relationships are the final category of visuals covered here, and they are often the workhorse of geometries because they include the popular scatterplot (Figures 1C and 1D) and other presentations of x- and y-coordinate data. The basic scatterplot remains very effective, and layering information by modifying point symbols, size, and color are good ways to highlight additional messages without taking away from the scatterplot. It is worth mentioning here that scatterplots often develop into line geometries (Figure1D), and while this can be a good thing, presenting raw data and inferential statistical models are two different messages that need to be distinguished (see Data and Models Are Different Things).

Finally, it is almost always recommended to show the data.7 Even if a geometry might be the focus of the figure, data can usually be added and displayed in a way that does not detract from the geometry but instead provides the context for the geometry (e.g., Figures 1D and 1E). The data are often at the core of the message, yet in figures the data are often ignored on account of their simplicity.

Principle #4 Colors Always Mean Something

The use of color in visualization can be incredibly powerful, and there is rarely a reason not to use color. Even if authors do not wish to pay for color figures in print, most journals still permit free color figures in digital formats. In a large study20 of what makes visualizations memorable, colorful visualizations were reported as having a higher memorability score, and that seven or more colors are best. Although some of the visuals in this study were photographs, other studies21 also document the effectiveness of colors.

In today's digital environment, color is cheap. This is overwhelmingly a good thing, but also comes with the risk of colors being applied without intention. Black-and-white visuals were more accepted decades ago when hard copies of papers were more common and color printing represented a large cost. Now, however, the vast majority of readers view scientific papers on an electronic screen where color is free. For those who still print documents, color printing can be done relatively cheaply in comparison with some years ago.

Color represents information, whether in a direct and obvious way, or in an indirect and subtle way. A direct example of using color may be in maps where water is blue and land is green or brown. However, the vast majority of (non-mapping) visualizations use color in one of three schemes: sequential, diverging, or qualitative. Sequential color schemes are those that range from light to dark typically in one or two (related) hues and are often applied to convey increasing values for increasing darkness (Figures 1B and 1F). Diverging color schemes are those that have two sequential schemes that represent two extremes, often with a white or neutral color in the middle (Figure1E). A classic example of a diverging color scheme is the red to blue hues applied to jurisdictions in order to show voting preference in a two-party political system. Finally, qualitative color schemes are found when the intensity of the color is not of primary importance, but rather the objective is to use different and otherwise unrelated colors to convey qualitative group differences (Figures 1A and 1G).

While it is recommended to use color and capture the power that colors convey, there exist some technical recommendations. First, it is always recommended to design color figures that work effectively in both color and black-and-white formats (Figures 1B and 1F). In other words, whenever possible, use color that can be converted to an effective grayscale such that no information is lost in the conversion. Along with this approach, colors can be combined with symbols, line types, and other design elements to share the same information that the color was sharing. It is also good practice to use color schemes that are effective for colorblind readers (Figures 1A and 1E). Excellent resources, such as ColorBrewer,22 exist to help in selecting color schemes based on colorblind criteria. Finally, color transparency is another powerful tool, much like a volume knob for color (Figures 1D and 1E). Not all colors have to be used at full value, and when not part of a sequential or diverging color scheme—and especially when a figure has more than one colored geometry—it can be very effective to increase the transparency such that the information of the color is retained but it is not visually overwhelming or outcompeting other design elements. Color will often be the first visual information a reader gets, and with this knowledge color should be strategically used to amplify your visual message.

Principle #5 Include Uncertainty

Not only is uncertainty an inherent part of understanding most systems, failure to include uncertainty in a visual can be misleading. There exist two primary challenges with including uncertainty in visuals: failure to include uncertainty and misrepresentation (or misinterpretation) of uncertainty.

Uncertainty is often not included in figures and, therefore, part of the statistical message is left out—possibly calling into question other parts of the statistical message, such as inference on the mean. Including uncertainty is typically easy in most software programs, and can take the form of common geometries such as error bars and shaded intervals (polygons), among other features.15 Another way to approach visualizing uncertainty is whether it is included implicitly into the existing geometries, such as in a box plot (Figure1E) or distribution (Figures 1B and 1G), or whether it is included explicitly as an additional geometry, such as an error bar or shaded region (Figure1D).

Representing uncertainty is often a challenge.23 Standard deviation, standard error, confidence intervals, and credible intervals are all common metrics of uncertainty, but each represents a different measure. Expressing uncertainty requires that readers be familiar with metrics of uncertainty and their interpretation; however, it is also the responsibility of the figure author to adopt the most appropriate measure of uncertainty. For instance, standard deviation is based on the spread of the data and therefore shares information about the entire population, including the range in which we might expect new values. On the other hand, standard error is a measure of the uncertainty in the mean (or some other estimate) and is strongly influenced by sample size—namely, standard error decreases with increasing sample size. Confidence intervals are primarily for displaying the reliability of a measurement. Credible intervals, almost exclusively associated with Bayesian methods, are typically built off distributions and have probabilistic interpretations.

Expressing uncertainty is important, but it is also important to interpret the correct message. Krzywinski and Altman23 directly address a common misconception: “a gap between (error) bars does not ensure significance, nor does overlap rule it out—it depends on the type of bar.” This is a good reminder to be very clear not only in stating what type of uncertainty you are sharing, but what the interpretation is. Others16 even go so far as to recommend that standard error not be used because it does not provide clear information about standard errors of differences among means. One recommendation to go along with expressing uncertainty is, if possible, to show the data (see Use an Effective Geometry and Show Data). Particularly when the sample size is low, showing a reader where the data occur can help avoid misinterpretations of uncertainty.

Principle #6 Panel, when Possible (Small Multiples)

A particularly effective visual approach is to repeat a figure to highlight differences. This approach is often called small multiples,7 and the technique may be referred to as paneling or faceting (Figure1G). The strategy behind small multiples is that because many of the design elements are the same—for example, the axes, axes scales, and geometry are often the same—the differences in the data are easier to show. In other words, each panel represents a change in one variable, which is commonly a time step, a group, or some other factor. The objective of small multiples is to make the data inevitably comparable,7 and effective small multiples always accomplish these comparisons.

Principle #7 Data and Models Are Different Things

Plotted information typically takes the form of raw data (e.g., scatterplot), summarized data (e.g., box plot), or an inferential statistic (e.g., fitted regression line; Figure1D). Raw data and summarized data are often relatively straightforward; however, a plotted model may require more explanation for a reader to be able to fully reproduce the work. Certainly any model in a study should be reported in a complete way that ensures reproducibility. However, any visual of a model should be explained in the figure caption or referenced elsewhere in the document so that a reader can find the complete details on what the model visual is representing. Although it happens, it is not acceptable practice to show a fitted model or other model results in a figure if the reader cannot backtrack the model details. Simply because a model geometry can be added to a figure does not mean that it should be.

Principle #8 Simple Visuals, Detailed Captions

As important as it is to use high data-ink ratios, it is equally important to have detailed captions that fully explain everything in the figure. A study of figures in the Journal of American Medicine8 found that more than one-third of graphs were not self-explanatory. Captions should be standalone, which means that if the figure and caption were looked at independent from the rest of the study, the major point(s) could still be understood. Obviously not all figures can be completely standalone, as some statistical models and other procedures require more than a caption as explanation. However, the principle remains that captions should do all they can to explain the visualization and representations used. Captions should explain any geometries used; for instance, even in a simple scatterplot it should be stated that the black dots represent the data (Figures 1C–1E). Box plots also require descriptions of their geometry—it might be assumed what the features of a box plot are, yet not all box plot symbols are universal.

Principle #9 Consider an Infographic

It is unclear where a figure ends and an infographic begins; however, it is fair to say that figures tend to be focused on representing data and models, whereas infographics typically incorporate text, images, and other diagrammatic elements. Although it is not recommended to convert all figures to infographics, infographics were found20 to have the highest memorability score and that diagrams outperformed points, bars, lines, and tables in terms of memorability. Scientists might improve their overall information transfer if they consider an infographic where blending different pieces of information could be effective. Also, an infographic of a study might be more effective outside of a peer-reviewed publication and in an oral or poster presentation where a visual needs to include more elements of the study but with less technical information.

Even if infographics are not adopted in most cases, technical visuals often still benefit from some text or other annotations.16 Tufte's works7^,24 provide great examples of bringing together textual, visual, and quantitative information into effective visualizations. However, as figures move in the direction of infographics, it remains important to keep chart junk and other non-essential visual elements out of the design.

Principle #10 Get an Opinion

Although there may be principles and theories about effective data visualization, the reality is that the most effective visuals are the ones with which readers connect. Therefore, figure authors are encouraged to seek external reviews of their figures. So often when writing a study, the figures are quickly made, and even if thoughtfully made they are not subject to objective, outside review. Having one or more colleagues or people external to the study review figures will often provide useful feedback on what readers perceive, and therefore what is effective or ineffective in a visual. It is also recommended to have outside colleagues review only the figures. Not only might this please your colleague reviewers (because figure reviews require substantially less time than full document reviews), but it also allows them to provide feedback purely on the figures as they will not have the document text to fill in any uncertainties left by the visuals.

Thanks go to the numerous students with whom I have had fun, creative, and productive conversations about displaying information. Danielle DiIullo was extremely helpful in technical advice on software. Finally, Ron McKernan provided guidance on several principles.

•

Steve Midway is an assistant professor in the Department of Oceanography and Coastal Sciences at Louisiana State University. His work broadly lies in fisheries ecology and how sound science can be applied to management and conservation issues. He teaches a number of quantitative courses in ecology, all of which include data visualization.

1. Stirling P. Power lines. NZ Listener. 1987:13–15. [Google Scholar]

2. Fleming N.D., Mills C. Not another inventory, rather a catalyst for reflection. To Improve the Academy. 1992;11:137–155. [Google Scholar]

3. Lane S., Karatsolis A., Bui L. Graphical abstracts: a taxonomy and critique of an emerging genre. In: Gossett K., editor. Proceedings of the 33rd Annual International Conference on the Design of Communication. 2015. [CrossRef] [Google Scholar]

4. Csuri C. Computer graphics and art. Proc. IEEE. 1974;62:503–515. [Google Scholar]

5. Wang G., Gregory J., Cheng X., Yao Y. Cover stories: an emerging aesthetic of prestige science. Public Underst. Sci. 2017;26:925–936. [PubMed] [Google Scholar]

6. Cleveland W.S. Graphs in scientific publications. Am. Stat. 1984;38:261–269. [Google Scholar]

7. Tufte E.R. vol. 2. Graphics Press; 2001. (The Visual Display of Quantitative Information). [Google Scholar]

8. Cooper R.J., Schriger D.L., Close R.J. Graphical literacy: the quality of graphs in a large-circulation journal. Ann. Emerg. Med. 2002;40:317–322. [PubMed] [Google Scholar]

9. Weissgerber T.L., Milic N.M., Winham S.J., Garovic V.D. Beyond bar and line graphs: time for a new data presentation paradigm. PLoS Biol. 2015;13:e1002128. [PMC free article] [PubMed] [Google Scholar]

10. Nature Communications . Macmillan Publishers Ltd.; 2015. Nature Collections: Visual Strategies for Biological Data, the Collected Points of View (2010–2015) [Google Scholar]

11. Wilkinson L. Springer Science & Business Media; 2013. The Grammar of Graphics. [Google Scholar]

12. Wickham H. Springer; 2016. ggplot2: Elegant Graphics for Data Analysis. [Google Scholar]

13. R Core Team . R Foundation for Statistical Computing; 2020. R: A Language and Environment for Statistical Computing. [Google Scholar]

14. O’Donoghue S.I., Baldi B.F., Clark S.J., Darling A.E., Hogan J.M., Kaur S., Maier-Hein L., McCarthy D.J., Moore W.J., Stenau E. Visualization of biomedical data. Annu. Rev. Biomed. Data Sci. 2018;1:275–304. [Google Scholar]

15. Wilke C.O. O’Reilly Media; 2019. Fundamentals of Data Visualization: A Primer on Making Informative and Compelling Figures. [Google Scholar]

16. Lane D.M., Sándor A. Designing better graphs by including distributional information and integrating words, numbers, and images. Psychol. Methods. 2009;14:239–257. [PubMed] [Google Scholar]

17. Streit M., Gehlenborg N. Bar charts and box plots. Nat. Methods. 2014;11:117. [PubMed] [Google Scholar]

18. Annesley T.M. Bars and pies make better desserts than figures. Clin. Chem. 2010;56:1394–1400. [PubMed] [Google Scholar]

19. Tukey J.W. vol. 2. Addison-Wesley; 1977. (Exploratory Data Analysis). [Google Scholar]

20. Borkin M.A., Vo A.A., Bylinskii Z., Isola P., Sunkavalli S., Oliva A., Pfister H. What makes a visualization memorable? IEEE Trans. Vis. Comput. Graph. 2013;19:2306–2315. [PubMed] [Google Scholar]

21. Spence I., Kutlesa N., Rose D.L. Using color to code quantity in spatial displays. J.Exp. Psychol. Appl. 1999;5:393–412. [Google Scholar]

22. Harrower M., Brewer C.A. Colorbrewer.org: an online tool for selecting colour schemes for maps. Cartogr. J. 2003;40:27–37. [Google Scholar]

23. Krzywinski M., Altman N.S. Points of significance: error bars. Nat. Methods. 2013;10:921–922. [PubMed] [Google Scholar]

24. Tufte E.R. Graphics Press; 2006. Beautiful Evidence. [Google Scholar]

25. Gelman A., Pasarica C., Dodhia R. Let’s practice what we preach: turning tables into graphs. Am. Stat. 2002;56:121–130. [Google Scholar]

26. Spence I., Lewandowsky S. Displaying proportions and percentages. Appl. Cogn. Psychol. 1991;5:61–77. [Google Scholar]