Estimating Correlations from Scatterplots

Results indicate that estimates of correlations from scatterplots are partly based on perceptual processes that are influenced by visual properties of the display and are unrelated to the cognitive structures created by formal statistical training.

Abstract

Previous attempts to establish the function relating intuitive estimates of correlations from scatterplots to accepted statistical measures have led to unsatisfying results. In this study two experiments dealt with the effects of the statistical training of the viewer and various characteristics of the display on estimates. Statistical knowledge was related to higher estimates of correlations and the use of a wider range of values, but people with and without statistical knowledge were equally affected by the type of dispersion of the point cloud, the mere display of the regression line, and the slope of the regression line. Results indicate that estimates of correlations from scatterplots are partly based on perceptual processes that are influenced by visual properties of the display and are unrelated to the cognitive structures created by formal statistical training.