Choose honest charts

Chart Selection & Visual Encoding

Use this skill when choosing chart types, mapping data attributes to visual channels, and designing scales, axes, and legends.

Perceptual Accuracy & Task-Driven Encoding (Munzner & Cleveland)

1. Match the Chart to the Analytical Task and Data Type:
   • Trend & rate of change over ordered time: Line chart, step chart (for discrete state/price changes), or horizon/area chart.
   • Magnitude comparison across categories: Horizontal or vertical bar chart, Cleveland dot plot.
   • Distribution & outliers: Histogram, box/violin plot, strip/beeswarm plot, or quantile band (never hide a bimodal distribution behind a single mean bar).
   • Part-to-whole composition: Stacked bar (for a few segments), 100% stacked bar, marimekko, or treemap (avoid multi-slice pie/donut charts when $>3\text{–}4$ categories or close values must be compared).
   • Correlation & joint variation: Scatter plot, hexbin/contour density plot, parallel coordinates, or heatmap.
2. Respect Cleveland & McGill's Perceptual Hierarchy:
   • Encode primary quantitative comparisons using position along a common aligned scale first, then length, before falling back to area, angle, or color saturation.
   • Never use 3D extrusions, perspective distortion, or decorative gauges that falsify perceived proportions.

Honest Scales, Baselines & Axes (Tufte)

• Mandatory Zero Baseline for Length/Area Encodings:
  • Bar and filled area charts encode magnitude by length/area and must start at 0. If a narrow non-zero range is needed to inspect small variations, use a line or dot plot with clearly labeled axis bounds.
• Avoid Deceptive Dual Y-Axes:
  • Plotting two metrics on independent left/right Y-axes lets arbitrary scale stretching manufacture fake visual correlation. Prefer indexed small multiples (% change) or vertically stacked linked panels sharing an X-axis.
• Consistent Scales Across Small Multiples:
  • Lock side-by-side comparison panels to a shared scale by default; if independent scaling is toggled, mark the axis warning unmistakably.
• Explicit Non-Linear Scales:
  • Use log scales only across multiple orders of magnitude or multiplicative growth rates, and label ticks and axis titles clearly (log10 scale).

Checklist

• Does the chosen chart type directly answer the user's comparison, trend, distribution, or composition question?
• Do all bar and area charts start at a 0 baseline without arbitrary axis truncation?
• Have dual-axis distortions and 3D/area misproportions been avoided in favor of aligned common scales or linked panels?
• Are units, tick formatting, scale types (linear vs. log), and legends unambiguous?