Correlation Estimates as Perceptual Judgments

Correlation estimates from scatterplots were studied as an example for an intuitive decision task. Three experiments showed that subjective correlation estimates are based on geometric properties of the displays. People with different levels of statistical training were found to assess correlations from scatterplots in close accordance with the power function r_est = 1 - aX^b, where X is the mean of the geometrical distances between the points and the regression line or a similar central axis. Changes of the slope of the displayed point cloud and the introduction of outliers affected estimates as predicted from the function. The study demonstrated that intuitive judgments in a complex domain are based on the perception of geometric features of the relevant information. By applying these findings, graphic designers can accurately predict how changes in a display will affect viewers' impressions.