Abstract
Previous work has shown that the perception of correlation in scatterplots can be characterized by two simple laws: a linear Fechner-like law for precision and a logarithmic Weber-like law for accuracy (Rensink & Baldridge, 2010). It also appears to be rapid, being largely complete within 100 ms of presentation (Rensink, 2011). This suggests that although correlation may be conveyed by a complex carrier, it nevertheless is−or at least, is based on−a relatively simple property.
To investigate the nature of the process involved, two sets of experiments tested whether different kinds of visual design influence correlation perception. The first set involved scatterplots with various styles of dot (or symbol). Precision was determined via the just noticeable difference in correlation for two side-by-side scatterplots. Accuracy was determined by direct estimation, using reference scatterplots having fixed upper and lower values, and a test scatterplot adjusted to have its apparent correlation be midway between them. The second set used similar methodology but a different carrier, with the vertical position carrying the second data dimension being replaced by a simple feature such as size. Such ‘augmented stripplots’ then conveyed correlation via the relation between horizontal position and size. Twelve observers were tested in each condition.
Results showed a surprising degree of invariance for scatterplot symbol: different sizes, colors, and even shapes had little effect on precision or accuracy. This suggests that only the centers of the symbols are relevant, ruling out the involvement of simple operations such as blurring. In addition, there was also an interesting degree of invariance for carrier: accuracy and precision in augmented stripplots obeyed linear-logarithmic laws similar to those for scatterplots. These invariances suggest that correlation perception may be a general process that is both rapid and sophisticated.
Meeting abstract presented at VSS 2012