“They began three and a half centuries ago,” writes Gernsbacher (2018, p. 403). They can delight us or frighten us, teach us and confuse us, intimidate us or encourage us. They are the base unit of productivity and the currency of academic prestige and advancement. “They,” of course, are scientific journal articles.
The professional academic often fills a lot of different roles. We are researchers and educators, yes – but equally often we take on administrative roles. We make budgets; we approve expenses. We are employers, managers, and mentors.
And we are science communicators.
We are usually not formally trained as writers or as speakers, but these are central roles all the same. Writing articles is how we communicate our latest findings to our peers, our funders, and to the general public.
As psychologists, we might be expected to have some expertise in the area of efficient communication. After all, efficiency in communication is largely a matter of encoding information in such a way that it can be rapidly and flawlessly decoded – exactly the kinds of mental processes that we study. Most researchers, however, do not seem to apply such principles to their own writing, figures, or conference presentations.
In a recent article in the Psychonomic Bulletin & Review, Lei Yuan, Steve Haroz, and Steven Franconeri present an interesting new observation that might affect how you create your next graph. The paper starts off from the classic observation that certain visualizations are easier to interpret correctly than others – for example, expressing quantities as locations in a graph leads to more precise interpretations than encoding them as areas in a chart. This is illustrated in the figure below.
Data presented as extent (left), extent and position (center), and position only (right). A classical observation, due to Cleveland and McGill (1984), is that perceptual precision is greater when position is used to encode magnitude. (Image credit: Steve Haroz.)
Yuan and colleagues extended this question to comparisons of groups of markers, rather than one-versus-one comparisons as in the figure above. In one of their experiments, participants performed single-value comparisons (1 vs. 1 unit) as well as multi-value comparisons (2 vs. 2 or 6 vs. 6 units), with stimuli that were either misaligned bar graphs (length only), aligned bar graphs (length and position), or dot plots (position only). The task instructions were relatively straightforward:
(You can try the task for yourself here.)
Consistent with the received wisdom, in the one-versus-one comparisons the misaligned bar graphs led to worse performance: participants were less able to pick out the longer bar from a one-versus-one comparison. There was no appreciable difference between the normal bar graphs and the dot graphs. However, the difference between misaligned bars and the other two modalities appeared to disappear in the six-versus-six comparison task.
In a follow-up experiment, Yuan and colleagues sought to identify why performance suffered across multiple presentation modalities in multi-value comparisons. One hypothesis inspired by the first experiment held that the perception of average magnitude is driven by the total, not average, mass of a stimulus graph. That is, participants might be treating entire sets of bars as a single unit – integrating the entire area of the presented stimulus and failing to correct for the number of bars. The two sets of bars in the next figure are identical but for order, yet they have different shape envelopes and might be perceived differently:
To put this hypothesis to the test, Yuan and colleagues constructed a pairwise comparison task in which the multi-value comparisons were intentionally unbalanced (e.g., 6 vs. 10 units). As a result, competing sets of stimuli had different spatial extents independently of any difference in the average length per bar or position per dot.
As predicted, participant performance suffered when comparing groups of data points with different set sizes – both the normal and the misaligned bar graphs led to poorer performance than the dot graph. The irrelevant spatial extent of the stimulus intrudes upon the participant’s magnitude judgments.
As the authors point out, findings like these have practical implications: for the purposes of scientific communication, dot plots are often better representations of group data; and if the mean is important it is probably a good idea to represent it explicitly in the graph.
The contributions of cognitive science to scientific communication are not limited to the construction of figures. The American Statistical Association’s recent statements on the use of p-values are essentially human-factors arguments: If a method is routinely misused, then it stands to be improved, replaced, or forgotten. Indeed, statisticians like Sander Greenland make psychological arguments in their evaluation of tools and procedures. In my personal experience, psychologists and cognitive scientists tend to underestimate the usefulness of our field and expertise.
As former APA President E. G. Boring would have had it, “Ideally, it might be argued, the psychologist is a superior being, for over all other scientists he has the advantage of being a psychologist. He alone, the argument would continue, knows the human mind without which there could be no science.” (Boring, 1929, p. 97). Or statistics, or graphs.
Psychonomics article featured in this post:
Yuan, L., Haroz, S., & Franconeri, S. (2019). Perceptual proxies for extracting averages in data visualizations. Psychonomic Bulletin & Review, 26, 669-676. DOI: 10.3758/s13423-018-1525-7.