← All guides

Communicate · Sharing the work

Make graphs that explain results

Choose a display that matches the question, show observations honestly, and write captions that state the real takeaway.

Read
7 minutes
Updated
2026-09-22

Choose a graph from the question, not from the chart menu.

Match form to purpose

  • Compare distributions: dot plots, box plots, violin plots, or histograms—often with individual observations.
  • Compare a small number of estimates: points with intervals; bars only when length from a meaningful zero is the message.
  • Show change over time: line plots with the actual time scale and missing periods visible.
  • Show a relationship: scatterplots, optionally with a justified fitted model and uncertainty.
  • Show composition: stacked displays only when part-to-whole comparison is genuinely important.
  • Show spatial structure: maps or images with a scale, legend, and documented processing.

A pie chart is not the automatic choice for percentages. A bar chart is not the automatic choice for every comparison.

Hard to trust

9896Thing AThing B

Truncated scale, decorative depth, no units, no observations, no uncertainty.

Evidence first

BaselineNew method Error rate (%)

Labeled units, honest scale, observations, summary, uncertainty, and sample size in the caption.

Make the evidence inspectable

Label axes, units, conditions, and sample size. Show actual observations when practical. Define every line, band, error bar, normalization, and transformation. Use the same scale across plots that invite direct comparison.

Do not use 3D perspective, broken axes, cropped ranges, or area/volume effects to exaggerate differences. A nonzero axis can be appropriate for continuous measurements when the visible range is clearly labeled and interpretation is not distorted; explain important choices.

Write a takeaway caption

Bad: “Figure 2. Results.”

Better: “The motion-aware controller reduced median localization error by 24% versus the reactive baseline across 30 held-out simulations; points show runs and bars show 95% bootstrap confidence intervals.”

The caption should let a reader understand the claim, comparison, sample, and uncertainty without searching the paragraph above it.

Use color as reinforcement

Choose a restrained, color-vision-aware palette. Keep the same condition the same color across every figure. Add direct labels, symbols, line styles, or patterns so color is not the only way to distinguish groups.

For digital publication, provide a short text alternative and a longer description when the chart carries details not available nearby. The W3C complex-image guidance gives patterns for accessible chart descriptions.

Preserve the unflattering data

Show relevant failed trials, outliers, and exclusions according to rules defined before the result. If an observation is removed, state the reason and show whether the conclusion changes. Never select only favorable runs or smooth a line without disclosing the method.

Primary sources

Go deeper

Use these first-party references to check rules, study the method further, or adapt this guide to your field.