Writing · Beginner
Design honest figures and write complete legends
Design figures with clear encodings, intact evidence and complete legends.
10 min read · Oct 26, 2026
Key points
- Choose the plot from the question.
- Label units, observations and uncertainty.
- Preserve original image evidence.
Contents
Design figures with clear encodings, intact evidence and complete legends.
A figure is part of the evidence in a paper. Its job is to help readers inspect a comparison, understand a procedure or evaluate a claim. A technically correct analysis can still be difficult to judge if the figure hides individual observations, uses unexplained error bars or makes different groups appear comparable when they were measured differently. Good figure design begins before choosing colours.
By the end of this tutorial, you should be able to choose an appropriate display, explain what every visual element represents, write a complete legend and check the exported figure. The same reasoning applies to a school project, a laboratory study, a social survey, an engineering evaluation and a review article.
Start with a question the figure must answer
Write a provisional sentence such as “Does the change vary between participants?” or “How do measurement error and energy consumption trade off?” Then list the evidence needed to answer it. The first question needs matched observations or changes for each participant. The second needs both performance measures for the same system configurations. A chart showing only group means would not answer either question fully.
A visual encoding is the rule that maps a value to position, length, colour, shape or another visible property. A panel is one component of a larger figure. Decide what one mark represents before deciding how attractive it should look. One point might be one participant, one independent culture, one simulation scenario or one estimated coefficient. Those meanings are not interchangeable.
Use this planning sequence:
- State the comparison or pattern of interest.
- Identify the unit represented by each observation.
- Decide whether readers need individual values, a distribution, an estimate with uncertainty, or a diagram.
- Choose axes and visual encodings that reveal that information.
- Draft the legend before polishing the appearance.
If the legend is difficult to write, the figure may contain an unresolved design or analysis problem. For example, “each point is a replicate” is insufficient if some points are repeated readings of one specimen and others are separate specimens.
Select the display by its scientific purpose
| Reader's question | Useful starting display | Essential information |
|---|---|---|
| How do two numeric variables vary together? | Scatter plot | Unit, axes, overlapping values |
| What changed in the same units? | Paired plot or difference plot | Matching and direction of subtraction |
| How variable are the observations? | Dot, box, histogram or violin plot | Sample size and summary conventions |
| How large is an estimated effect? | Point and interval plot | Effect scale, reference and interval type |
| What is the composition of a sample? | Categorical bar chart | Counts, categories and denominator |
| How was the study conducted? | Procedure diagram | Sequence, exclusions and decisions |
| How might a mechanism work? | Conceptual diagram | Which relationships are established or proposed |
A diagram can explain an assumption without providing evidence for it. In a review of learning theory, arrows between motivation, practice and performance might summarise a proposed explanation. The legend should state whether the arrows represent tested causal effects, associations or hypotheses. Cite the sources behind substantive relationships rather than making every arrow look equally established.
Preserve the observations that matter
A summary can conceal scientifically important differences. Two treatments can have similar averages while one produces consistent outcomes and the other produces extreme successes and failures. A paired design can show improvement within nearly every participant even when the two group distributions overlap. Weissgerber and colleagues illustrate why plots of individual values and paired observations often reveal information hidden by bars of means. S402
Showing all observations is not always possible. Thousands of points can overlap, and individual records may reveal sensitive information. Options include transparency, bins, distributions, aggregate summaries and accessible underlying tables. Describe what the display suppresses or aggregates. A plot of classroom means may be appropriate, but call the points classrooms and state how many pupils contributed to each mean.
Make scale choices visible
Label every axis with a meaningful variable name and unit. “Response” is weaker than “Plant dry mass, g” or “Completion time, s”. A bar length usually requires a zero baseline because readers compare lengths from that baseline. A point and interval plot can use a restricted range when that range is clearly shown. The purpose is to reveal the relevant comparison without exaggerating it.
An axis transformation changes how distances are interpreted. On a base 10 logarithmic axis, equal distances represent equal ratios, such as 1 to 10 and 10 to 100. They do not represent equal arithmetic differences. Ordinary logarithms cannot display zero or negative measurements. State the transformation and any handling of such values; never quietly remove them.
Use consistent limits across panels intended for direct comparison. If different limits are necessary, make the difference conspicuous. Avoid two vertical axes unless there is a clear reason and an explicit explanation, because arbitrary scaling can make unrelated series appear to track each other closely.
Explain uncertainty rather than decorating with it
An error bar is not a universal statistical object. Standard deviation, or SD, describes spread among observations. Standard error, or SE, describes the estimated sampling variability of an estimator. A confidence interval expresses uncertainty about a parameter under a specified method and assumptions. A prediction interval addresses a future observation and generally includes variation that a mean confidence interval does not.
State the interval level and calculation. For repeated observations, the method must respect the dependency structure. A narrow interval calculated as though every sensor reading came from a separate device can be misleading even if the plot itself is beautifully drawn.
Avoid interpreting overlap of two separate group intervals as a general test of their difference. A paired comparison should normally show uncertainty for the paired difference itself. The estimate of the contrast answers the question more directly than readers trying to infer it from two error bars.
Use colour and labels accessibly
Use colour to communicate information, not simply to fill space. A sequential palette progresses through an ordered magnitude. A diverging palette distinguishes values on either side of a meaningful centre. A categorical palette separates groups without suggesting numeric order. Colour gradients with uneven perceptual steps can create apparent boundaries absent from the data, as discussed by Crameri and colleagues. S400
Support colour with shape, line style, direct labels or position. Check the figure at its intended size and in greyscale. A legend that requires distinguishing six nearly identical colours may fail on a small screen even for readers with typical colour vision. Use plain names consistently: if the Methods calls a condition “low light”, do not label the figure “Group 2”.
Treat experimental images as measurements
Keep original image files and document transformations. Show scale bars with units for spatial images, identify crops and composites, and make comparable adjustments across images intended for comparison. Do not selectively erase a band, remove background features or enhance one region to strengthen a result. PLOS's figure guidance sets out these image integrity principles and its own requirements for underlying blot and gel files. Those publication requirements should be checked against the destination journal. S18
An illustration of a cell is artwork. A microscope image is an observation. Label them accordingly. Generated imagery must not replace missing experimental content. For an engineering photograph, record which prototype version is shown and provide a scale or dimensions where size matters. For a social science photograph, consider consent and identifiability before including it.
Build a legend readers can use independently
A legend should identify the subject, explain panels, define marks and summaries, give relevant sample sizes, state uncertainty, identify key analysis choices and expand abbreviations. Detailed procedures belong in Methods, but a reader should not need to search there to discover what a point or error bar means.
Consider this weak legend for the supplied synthetic controller comparison:
Figure 1. B works better than A. Error bars show uncertainty. A significant difference was found.
It gives no unit, sample size, pairing, effect direction or interval definition. It also makes an empirical performance claim from teaching values. A stronger legend is:
Figure 1. Positioning error in twelve synthetic layout cases. In the paired panel, each line joins controller A and controller B values from the same case; the vertical axis gives error in millimetres. The estimation panel shows the mean within case difference, defined as A minus B, with a t based 95% confidence interval using eleven degrees of freedom. The mean difference is 3.58 mm and the interval is 2.95 to 4.22 mm. Positive differences indicate lower error for B. All values were constructed for teaching and provide no validation of an actual controller.
The revised version lets a reader recover the comparison and its limitations. Notice that it does not require an asterisk: the estimate and interval carry the numerical information.
For a hypothetical survey figure, the corresponding legend might state that bars represent the percentage of respondents selecting each concern, that multiple responses were permitted, and that the denominator excludes missing answers. For a hypothetical cell culture figure, it should distinguish cultures from cells and explain whether points show culture means or individual cells.
Write useful alternative text and export carefully
Alternative text supports readers who cannot use the image directly. For a simple figure, state its purpose and essential pattern. For a complex figure, provide a short identification plus a longer explanation or accessible data table nearby. W3C guidance describes this approach for complex images. S401
For the controller example, useful short text is “Paired synthetic controller errors; every B value is lower than its corresponding A value. The mean A minus B difference is 3.58 mm.” A longer description can give the interval and explain the two panels. Do not rely on “blue dots are above red dots” when the scientific meaning is available.
Retain an editable source and export for the actual destination. Raster resolution depends on the final display dimensions; increasing a resolution setting cannot recreate detail absent from the original. Check font size, line width, symbols, cropping and panel labels after export. A readable source file can become unreadable when reduced to a narrow website column.
Practice and model answer
Task: A figure compares a control and treatment using 120 cells from three cultures per condition. The legend says “n = 120, mean ± SE”. Identify what must be resolved before publication.
Model answer: The legend must specify whether 120 refers to each condition or the total, how cells are distributed among cultures, what received treatment and which units are independent. If culture is the experimental unit, treating cells as independent for the SE is inappropriate. Show culture level information or use a justified hierarchical analysis. Define the displayed summary and uncertainty, state the independent sample size, and describe the cell sampling rule. Changing the legend alone cannot repair an invalid calculation.
Use Template 11 to draft the legend and Template 15 to audit the final figure. The final question is practical: can another reader tell what was measured, how it was summarised and how far the claim can reasonably extend?