Your assistant hands you a colorful bar chart of average sales per square foot by borough. Naming three concrete flaws, and the fix for each, is a repeatable exam point.
Three ways a graph goes wrong. Kieran Healy frames bad graphs as failures of taste, of substance, or of perception. A taste problem is chartjunk: gratuitous 3D, gradients, and clip-art that add ink but no meaning. A substantive problem misleads about the data, like a y-axis that starts above zero and exaggerates differences. A perceptual problem picks an encoding the human eye reads badly, like a pie with a dozen near-equal slices.
KEY: When asked to critique a graph, sort your complaints into these buckets. Each distinct flaw plus a fix is worth its own point.
How perception ranks encodings. People judge some visual channels more accurately than others. Position along a common scale is best, then length, then angle and slope, then area, and color and shading are worst.
Common mistakes
- Critiquing the model, not the graph. When asked for graph issues, stay on presentation: axes, ordering, encoding, labels. Modeling comments earn no credit here.
- Recommending stacked histograms. For a numeric distribution across categories, stacking obscures each group. Use split or faceted histograms or boxplots instead.
- Leaving the y-axis truncated. A baseline above $0 exaggerates differences. Note it as a distortion and recommend starting at zero.
Bottom line
- Good graphs fail three ways: bad taste (chartjunk), bad data (misleading), and bad perception (hard-to-decode encodings).
- The eye decodes position and length accurately; it decodes angle, area, and color poorly, so prefer bars and dots over pies and 3D.
- ggplot2 builds a plot in layers: data, aesthetic mappings (aes), geoms, scales, facets, coordinates.
- Maximize the data-to-ink ratio: strip gridlines, 3D, shadows, and decoration that carry no information.
Exam shortcut
For a graph-critique subtask, produce distinct issues from three buckets (misleading data like a truncated axis, chartjunk like 3D, and hard-to-decode encodings like color for magnitude), then give one fix each. For bivariate pairings, memorize the map: categorical-categorical is bar, categorical-numeric is split boxplot or histogram, numeric-numeric is scatterplot; never answer stacked histogram for comparing distributions.
The full lesson (about 1,529 words, 10 min read) adds 2 worked examples, all 6 common mistakes, a self-check, free in the app.
Learning objectives
- 2d
Browse all free Exam PA lessons or jump into free Exam PA practice questions.