Abstract
Several high profile editorials and articles have advocated replacing the ubiquitous, but not especially informative, bar plot with dot plots or box plots which show the data or some summary of the data distribution. I argue that the biggest deficiency of bar plots (and other mean-and-error plots) is that they fail to directly communicate modeled effects and uncertainty, and dot and box plots do not address this deficiency. A Harrell plot combines a forest plot of modeled treatment effects and uncertainty, a dot plot of raw data, and a box plot of the distribution of the raw data into a single, two-part graph. A Harrell plot encourages best practices such as communication of the distribution of the data and a focus on effect size and uncertainty, while discouraging poor practices such as hiding distributions and focusing on p-values.
Footnotes
↵1 walker{at}maine.edu