Abstract
We think data is definitive, but our perception of it contains bias from expectations and motivations. For example, when Democrats and Republicans view the same depiction of global temperature trends, Democrats see an increasing trend, while Republicans see overall flatness. Could prior beliefs bias our perception of relations depicted in visualized data? We empirically examine how prior beliefs influence correlation estimations in scatterplots. We recruited 295 adults from Amazon’s Mechanical Turk and trained them to extract correlations from scatterplots. Participants viewed four no-context scatterplots (axes ‘X’ and ‘Y’) and the same four scatterplots, but with context (axes with real-world variable pairs). Participants also rated their agreement with statements describing the variables’ relationship, which were used to measure prior beliefs about the pairs. People who more strongly believed a relationship depicted with the real-world variables estimated the correlation to be stronger in the plot with context than in the plot without context. People who did not believe in the relationship estimated the correlation to be weaker for the context plots than for the no-context plots. These results suggest that real-world context around a scatterplot can generally bias correlation estimation, and for belief-triggering topics, the amount of bias can depend on the strength of the belief.