brycewang-stanford/figures

Figure Designer

Heritage and scope

This is an original Open Science Skills workflow for figure production in social-science manuscripts. It is general — apply to any figure type (line, bar, point, density, map, network, small multiples). For figures whose interpretation depends on method-specific standards, also consult the relevant sibling skill (conjoint-design, conjoint-diagnostics, list-experiment, topic-modeling, text-classification, vlm-ocr-pipeline). For end-stage QA on a finished figure set, hand off to figure-table-audit.

A good figure earns its place in the manuscript: it makes a single comparison legible, it can be read without the surrounding text, and it can be regenerated from a script. If a figure cannot do all three, it is not yet ready.

Instructions

1. State the comparison before drawing

Write down, in one sentence, what the figure is supposed to let the reader see. Examples:

• "Treatment effect is larger among low-information respondents than among high-information respondents."
• "Vote share for Party A declines monotonically with district urbanization."
• "Topic 7 prevalence rises sharply after the 2024 election."

If you cannot state the comparison in one sentence, the figure has too many goals — split it into multiple panels or multiple figures.

2. Choose the chart type from the comparison

Match the geometry to the comparison, not to the data type:

• Comparing magnitudes across a small number of categories: bar chart (or dot/lollipop for many categories).
• Tracking a quantity over a continuous variable: line chart for ordered/continuous x; point + ribbon for uncertainty.
• Showing distribution: histogram (single), density (compare a few), boxplot/violin (compare many).
• Showing relationship between two continuous variables: scatter, with trend line + uncertainty band if a model is implied.
• Showing effect estimates: coefficient plot (point + CI), preferred over tables for visual comparison across models.
• Showing many small comparisons: small multiples (faceting) on a shared scale, not multiple separate figures.

Avoid pie charts, 3D anything, dual-axis, and donut charts in academic figures.

3. Pick scales and axes deliberately

• Always label both axes with the quantity AND the units. "Income" is wrong; "Household income (USD, 2020)" is right.
• For percentages, write 0.00–1.00 or 0%–100% consistently — do not mix within a figure.
• Use a log scale when the data span more than two orders of magnitude OR when a multiplicative effect is the substantive story; label the scale as log.
• Anchor axes at zero ONLY when the zero is meaningful for the comparison (counts, proportions). For deviations, differences, and z-scores, do not force a zero baseline.
• Use the same y-axis range across panels of small multiples unless the claim is explicitly about within-panel scale.
• Reverse the y-axis only when the substantive direction demands it (e.g., rank where 1 is "best") and label the reversal in the caption.

4. Choose color with intent

• Sequential (light → dark): for ordered/quantitative variables (income brackets, time, density).
• Diverging (red ↔ blue, etc.): for variables with a meaningful midpoint (effect sign, deviation from baseline).
• Categorical: for unordered groups; use a colorblind-safe palette (Okabe-Ito, viridis discrete, ColorBrewer Set2/Dark2).
• Test grayscale: convert the figure and check that the comparison still reads. If it does not, add shape, line type, or direct labels.
• Do not encode the same variable in both color and shape unless reinforcing for accessibility — that is fine; just keep the legend honest.
• Default rainbow / jet ramps are a publication smell; replace them.

5. Match legend order to the visual order in the chart

Legends are read alongside the plot, so the legend order must mirror the data's visual order — readers should never have to scan back and forth to decode a series:

• For lines, areas, or stacked elements arranged top-to-bottom in the plot at the rightmost x-value, list legend entries in the same top-to-bottom order.
• For bars or categories arranged left-to-right, use a horizontal legend listed left-to-right in the same order.
• For panels (small multiples), follow the panel grid: rows top-to-bottom, columns left-to-right.
• Do NOT alphabetize the legend when the data have a natural order (time, magnitude, treatment intensity, ordinal scale). Alphabetical legends are a frequent cause of misreading.
• Prefer direct labeling at the line end / bar end when there are five or fewer series — a labelled line beats any legend.
• When a legend is unavoidable, place it where it does not occlude data, and match colors and line styles between figure and legend exactly.

The rule of thumb: the eye should be able to walk from chart to legend in the same direction it reads. Top-to-bottom on the chart maps to top-to-bottom in a vertical legend, and to left-to-right in a horizontal legend.

6. Write a self-contained caption

A reader who skims should be able to understand the figure from the caption alone. Include:

• What is plotted (the dependent quantity, the unit of analysis).
• The sample (N, source, time window).
• The uncertainty interval (e.g., "95% confidence intervals from cluster-robust standard errors").
• Panel meanings, if multi-panel.
• Any transformations or exclusions.
• The estimand or model family if the figure is showing model output (not just "Results").

Keep abbreviations defined and units explicit.

7. Make it reproducible

• Generate the figure from a script (R, Python, Stata, Julia) checked into the replication archive — never hand-edit final figures in Illustrator or Inkscape unless the manual step is documented.
• Save with the script: input data path, package versions, random seed (for jittered or sampled plots), output dimensions.
• Export vector (PDF, SVG) for plots with text and lines; export raster (PNG at ≥300 dpi) only for raster-native content (heatmaps with thousands of cells, photographs, maps with imagery).
• Use a single ggplot/matplotlib theme across the manuscript so figures are visually consistent.
• For multi-panel figures, build with patchwork/cowplot/gridExtra or matplotlib's subplots/gridspec — not by stitching exported PNGs in Word.

8. Accessibility and production sanity

• Axis label and tick text ≥ 8pt at print size.
• No reliance on color alone (pair with shape, line type, or direct labels).
• Alt text for journals that require it (a one-sentence description of what the figure shows).
• Consistent figure dimensions and fonts across the manuscript.
• Title text inside the plotting area is usually unnecessary — captions carry the title.

Output

When asked to design or revise a figure, produce:

# Figure Plan

Comparison: <one sentence>
Chart type: <type and why>
Geometry: <axes, scales, faceting>
Color encoding: <palette, what it encodes, accessibility check>
Legend / labeling: <direct label or legend; order matches visual order>
Caption draft: <self-contained>
Reproducibility: <script path, packages, output format and dimensions>
Open issues: <anything that needs author input — denominator choice, sample restriction, etc.>

When asked to produce code, default to a single ggplot2 (R) or matplotlib + seaborn (Python) script with the theme, palette, and figure dimensions explicit at the top.

Quality checks

• [ ] One-sentence comparison was stated before the chart was chosen.
• [ ] Chart geometry matches the comparison.
• [ ] Axes are labeled with quantity AND units.
• [ ] Color encoding is intentional and grayscale-tested.
• [ ] Legend order mirrors the visual order in the chart (no surprise alphabetical sort).
• [ ] Caption is self-contained: quantity, sample, uncertainty, panels, transformations.
• [ ] Figure is generated by a script, not hand-edited.
• [ ] Vector format used unless the content is raster-native.
• [ ] Final QA pass deferred to figure-table-audit once the figure set is stable.