Most bad figures are not badly drawn — they are badly chosen. A pie chart of a four-group comparison, a line plot of unpaired categories, a bar chart hiding a bimodal distribution: each answers a question the reader wasn't asking. This guide gives you a decision rule for choosing chart types that starts with one question — what relationship in the data am I claiming? — and walks through the strongest chart for each claim and when it can mislead.
Start With the Question, Not the Chart
Every figure makes a claim about structure: this group differs, these variables track each other, this quantity grows. Before opening any plotting library, finish this sentence: "The reader should see ______." The noun in that sentence picks your chart family:
- Distribution → violin or box plot
- Comparison between groups → bar chart (with caveats)
- Relationship between two continuous variables → scatter plot
- Composition of a whole → stacked bar
- Trend over an ordered axis → line plot
- Process or workflow → schematic
- Any of the above → show the uncertainty
Distributions → Violin or Box Plot
Use when the reader needs to see the shape, spread, and center of a variable — not just its mean. Is your data bimodal? Skewed? A bar chart of means will hide exactly that.
When it lies: box plots imply a symmetric, unimodal distribution that may not exist. A box plus whiskers over a bimodal sample describes a population that isn't there.
Pro tip: overlay the raw points (jittered) on top of the box or violin. Reviewers increasingly expect to see the actual observations, and overlaying points exposes outliers that summary statistics conceal.
Comparison → Bar Chart, Honestly
Use when comparing a summary statistic (usually a mean or count) across a handful of groups. Bars are fine for this — the trouble starts with what people put on them.
When it lies: the baseline. Bars encode value by length, so a truncated y-axis (starting at 95 instead of 0) visually doubles a 2% difference. It's the single most common form of accidental deception in research figures — and reviewers notice. Never truncate the axis on a bar chart; if small differences matter, show the distribution instead and let the spread make the argument.
Pro tip: order the bars by value, not alphabetically. An alphabetical order scrambles the pattern the reader is trying to see; a sorted order makes the effect size legible before anyone reads a label.
Relationship → Scatter Plot
Use when you're claiming two continuous variables covary — dose versus response, temperature versus rate. It is the most honest of all research data plots because every single observation is visible as its own mark.
When it lies: overplotting. Thousands of overlapping points can make a dense region look empty or, conversely, make noise look like a dense cluster. Outliers can hijack the visual impression of correlation.
Pro tip: add a fitted trend line with a confidence band, and state what the line is in the caption. A naked line invites readers to assume you fit something more rigorous than you did.
Composition → Stacked Bar
Use when showing how parts contribute to a whole, especially across conditions — cell-type proportions per sample, budget shares per grant year.
When it lies: stacked bars make every segment except the bottom one start from a different baseline, so middle segments are notoriously hard to compare across bars. Also avoid more than about five segments — beyond that, colors blur and the figure stops answering questions.
Pro tip: put the segment you care most about at the bottom, flush against the zero baseline. Its height can then be read directly and compared across every bar.
Trends → Line Plot
Use when the x-axis is ordered — time, temperature, concentration, dose — and consecutive points belong to a continuum. The connecting line says "these points are steps in a sequence," which is exactly why lines are wrong for unordered categories.
When it lies: two ways. Connecting points measured from different subjects pretends the intervening values exist. And dual y-axes can manufacture a visual correlation between two unrelated quantities by sheer axis scaling.
Pro tip: when multiple series come from independent replicates, plot the thin individual trajectories in grey underneath, and the group mean as a heavier line on top. It shows trend and variability together.
Process and Workflow → Schematic
Use when explaining method: study design, an experimental pipeline, an algorithm's steps. Charts show data; schematics show structure and sequence.
When it lies: it doesn't, statistically — but every box is a claim that this step matters, and over-detail wastes the reader's attention.
Pro tip: keep it to five to seven boxes in the main flow, with details pushed to labels or supplementary expansions. Readers should be able to trace the left-to-right narrative in one sweep.
Whatever You Choose, Show the Uncertainty
This is the non-negotiable one. A mean without a confidence interval, a line without a band, a bar without spread indicators — each presents a point estimate as if it were a fact. Show error bars or confidence intervals as standard, state in the caption what they represent (SD, SEM, or CI — reviewers check), and plot raw points where feasible. If a chart type cannot carry uncertainty, choose a different one.
A Decision Flow You Can Memorize
- What's the claim? Pick the noun: distribution, difference, relationship, composition, trend, or process.
- Distribution? Violin/box with raw points overlaid.
- Group comparison? Bars from zero, sorted by value — or distributions if the spread matters.
- Two continuous variables? Scatter with a fitted band.
- Parts of a whole? Stacked bar, key segment on the baseline.
- Ordered axis? Line plot, replicates shown under the mean.
- Method or pipeline? Schematic, five to seven boxes.
- Always: add uncertainty, state what it is, and design the result according to sound figure-design principles — our guide to research figure design principles picks up exactly there.
Once you know the chart you need, AI-assisted generation can scaffold the plot while you keep control of design decisions — see our AI scientific figure generation guide.
Key takeaways
- Choose the chart from the claim you're making, not from habit or your software's default.
- Bar charts must start at zero; truncated axes are the most common way research figures mislead.
- Overlay raw data wherever possible: points on boxes, replicates under means, every observation in scatters.
- Stacked bars hide middle segments — put your key category against the baseline, and cap segments at about five.
- Show uncertainty on every figure and state exactly what your error bars represent.
Frequently asked questions
When is a pie chart acceptable?
Almost never in a research manuscript. Readers compare angles poorly, and small slices become unlabelable. A bar chart conveys the same composition data with far greater accuracy. The rare exceptions are simple two-part splits in slides, better served by a donut anyway.
Violin plot or box plot — which should I use?
Box plots if your audience is conservative or the sample is small; violins when the shape of the distribution matters to your claim, such as bimodality or skew. With modest samples, show raw points regardless — box plus jittered points is the safest default.
Can I use a dual-axis chart in a paper?
It's legal but risky: dual axes let scaling choices manufacture or suppress apparent correlations. If two variables genuinely relate, a scatter plot shows that directly. If you must, label both axes clearly and let the statistics, not the scaling, carry the argument.
How do error bars differ from confidence intervals?
Standard deviation describes the spread of the data; standard error describes the precision of the mean estimate; a confidence interval gives a range of plausible values for the underlying parameter. They answer different questions. Whichever you plot, name it in the caption — an unlabeled error bar forces reviewers to guess.
The fastest way to internalize good chart choices is to see many examples. Browse the SciScroll discover feed to study how other researchers visualize data like yours, then generate your own figures with AI assistance — start free and upgrade via pricing if your lab needs more.