Although we would like to think of ourselves as independent readers and thoughtful evaluators of published data, it turns out that figure format heavily influences our perception of the data – even when the data points represented by the figure are identical. To look at it a different way, as authors, we can deliver our message and interpretation not only via text, but already in the figures. Research into the effect of a reader’s initial, “fast thinking” impressions only underlines the importance of thoughtfully building graphs congruent with the message of the paper.
In her studies, Prof. Barbara Tversky looks at the way people have been representing information visually for thousands of years. Through her research as well as that of others, it has become apparent, for example, that people intuitively represent time on horizontal axes. That such unspoken conventions have existed since the time of cavemen suggests the existence of intuitive principles that take effect during interpretation of visuals. To investigate these principles, Zacks and Tversky set out to probe the interpretation of bar versus line graphs by undergraduates at Stanford University. Specifically, the authors hypothesized that bar graphs will emphasize the difference between the different bars while line graphs will draw the interpreter’s attention to the trend as a function of the x-axis.
In a series of three experiments, the authors tested their hypothesis by first asking subjects to describe a graph they were shown and then asking subjects to draw a graph based on a description. While the data represented by the graphs were identical, subjects used words like “bigger” and “smaller” to describe bar graphs and “decreasing” and “increasing” to describe line graphs. This difference suggested that when looking at a bar graph, subjects focused on the difference between categories described by the bars. However, when looking at a line graph, subjects focused on the trend along the x-axis. In their second experiment, the authors labeled the graphs with “height” on the y-axis and either “male” and “female” or “10-” and “12-year olds” on the x-axis. Thus, the x-axis had either a discreet label (easily described using a “bigger/ smaller” comparison more appropriate for gender) or a continuous label (more suited to a “increasing/decreasing” trend description of change with age). The authors then mixed and matched labels and graph types creating consistent label-graph matches (gender for bars, age for line) and contradictory matches (gender for line, age for bars).
Statistical analysis showed that the impact of the graph type was much stronger than that of the label. That is, for contradictory matches, subjects were more likely to rely on graph type for description, which led to descriptions like “The more male the person is, the taller they are.” Finally, the third experiment was, in essence, a reversal of the second experiment. Subjects were asked to draw a graph based on a description (such as “Height for males is greater than for females.”) Once again, it was found that the word choice in the description influenced graph type choice more than the identity of the data did.
Combined, these experiments showed that the design of the graph exerts a powerful influence on the interpretation of the data. So how can we protect ourselves – and others – from the graph type bias? A number of studies have compared interpretation of graphs by experts and non-experts. Although there is still some disagreement in the field, it appears that at least some groups of experts are able to overcome graph type bias. The current hypothesis is that they do so by looking past the graph type; they consider trends and patterns instead of the absolute points. In other words, many years of training in thoughtful graph interpretation are likely to make your quick judgements about graphs more correct. But in the mean time, given the Dunning-Kruger effect, if you think you understand the figure, look at it another time – just in case.