People use graph in all walks of life and science as they can effectively and easily convey information through visual representations. However, nobody actually seriously look at the science of visualization until 232 years ago by a man named William Playfair(1786). There was some studies and debates on bar charts, pie charts, but graph design for data analysis and presentation is not scientific enough because most of the time people built graph based on intuition and rule of thumb There were not enough talks however on how to develop basic principles of graphic perception.
William Cleveland and Robert McGill(1984)however, has tried to set up a scientific foundation by exploring human graphical perception. The first step is “an identification of a set of elementary perceptual tasks when people extract quantitative information from graph.” (Cleveland and McGill,1984) The second is to order these tasks based on accuracy of how people perform. Based on these elements, they were able to draw a conclusion that in order to achieve more accuracy and increase the viewer’s ability to detect patterns and organize information, one needs to have more high-ranking tasks in one’s graph.

Here are the elementary perceptual tasks tested in their research: 1. Position along a common scale 2. Position along nonaligned scales 3. Length, direction, angle 4. Area 5. Volume, curvature 6. Shading, color saturation

Graphs which have high-ranking elementary perceptual tasks will have increased probabaility of identifying the correct pattern than those graph with lower-ranking elementary perceptual tasks. According to psychophysical theories and experimental results, position judgements are more accurate than length judgement and length judgement is more accurate than area and area judgement is better than volume.

A pie chart is not as good as a dot chart to achieve better accuracy. The reader would not be able to tell the ordering from the pie chart but can easily do so through the dot chart. This means a dot chart increases reader’s ability to detect patterns. Here is an example from Cleveland and McGill(1984)

Grouped dot chart can always replace divided bar chart. The reason is similar, length judgement is less accurate than position judgement and the readers can see the ordering of the items much more easily than a divided bar chart. For example here(Cleveland and McGill,1984) :

A curve difference chart is even worse than a divided bar chart as making a length judgement more inaccurate and difficult. You can replace it with a graph with task being judging position along a common scale. The remedy could be a aggregated line chart. For example here:(Cleveland and McGill,1984) Judging from shading/color saturation is at the bottom of the perceptual hierarchy. One should replace it with statistical map with located bars so that the elementary perception task is length, which has better rankings and helps deal with the contiguous clusters of states. The first graph is a color saturation graph and readers cannot easily tell which country has the highest rankings; The second graph is an improvement where color is replaced by length. Readers can easily organize the countries by rankings. Data source: CBRE EA

Despite the above analysis based on the role each elementary perception task plays, there are also other interesting analysis. Heer and Bostock(2010) replicate the old experiment and conducted new experiment using modern platform of Amazon’s Mechanic Turk. Their research was on on rectangular area perception and on chart size and gridline spacing. According to their study, the accuracy of rectangular area judgments matches that of circular area judgments and that rectangles with aspect ratio 1 has the worst performance. Heer and Bostock(2010) also made progress on chart size, their result shows that “gridlines should be spaced at least 8 pixels apart and increasing chart heights beyond 80 pixels provides little accuracy benefit on a 0-100 scale.” (Heer and Bostock,2010)

Drew Skau and Robert Kosara(2016) made some interesting research on pie chart and donut chart. Their results show that donut chart is no worse than pie chart as angle is not the primary or only factor in reading a pie chart. Thus removing the center does not make a difference.

Jeffrey Heer, Nicholas Kong, and Maneesh Agrawala researched into the the effects of chart Size and layering on the graphical perception of time series visualizations; They found that mirroring will not decrease accuracy;And that for each chart, there is an optimal chart size; For example, the optimal is a chart height of 25 pixels for both normal line charts and 1-band mirror charts.

There are many more interesting and informative research being done by scientists on how to best use visual representation to convey information.As Jeffrey Heer and Michael Bostock(2010) mentioned in their paper: “Assessing the impact of visual encodings on graphical perception enables designers to optimize their visualizations and is vital to the design of automatic presentation software.”. People need to optimize the accuracy of their visualizations by deploying the science of data visualization.

Reference

Cleveland, W. S., & McGill, R. (1984). Graphical perception: Theory, experimentation, and application to the development of graphical methods. Journal of the American statistical association, 79(387), 531-554.

Heer, J., & Bostock, M. (2010, April). Crowdsourcing graphical perception: using mechanical turk to assess visualization design. In Proceedings of the SIGCHI conference on human factors in computing systems (pp. 203-212). ACM.

Heer, J., Kong, N., & Agrawala, M. (2009, April). Sizing the horizon: the effects of chart size and layering on the graphical perception of time series visualizations. In Proceedings of the SIGCHI Conference on Human Factors in Computing Systems (pp. 1303-1312). ACM.

Skau, D., & Kosara, R. (2016, June). Arcs, angles, or areas: individual data encodings in pie and donut charts. In Computer Graphics Forum (Vol. 35, No. 3, pp. 121-130).