By Dr. Nedim Yel, Co-Founder & Lead Statistician · Updated
2 min read
Most figure advice is about aesthetics — fonts, colors, layout. That's worth getting right, but it's not what determines whether a figure is trustworthy. The rules below are the ones that actually affect whether a reader can trust what the figure is telling them, not just whether it looks polished.
Rules that actually matter
Always show uncertainty, not just a point estimate. A bar chart of means with no error bars or interval invites a reader to treat a possibly-noisy difference as solid fact.
Avoid the "dynamite plot" — a bar with a single error bar on top — for continuous data. It hides the actual shape of the distribution, and two groups with very different distributions can produce an identical-looking bar. A boxplot, violin plot, or raw jittered points shows what a bar and whisker can't.
Match the chart type to the data. Categorical comparisons: points or bars with intervals. Continuous relationships: a scatterplot, not a bar chart of binned averages that throws away information. Parts of a whole: a simple table or stacked bar, not a pie chart — pie charts are genuinely hard to compare precisely by eye once there are more than two or three slices.
Don't truncate a bar chart's y-axis without an unmistakable break marker. Starting the axis above zero exaggerates every difference in the chart and is one of the most common ways a completely accurate set of numbers ends up telling a misleading story.
Use vector formats — PDF, EPS, SVG — for anything headed to print. A raster PNG or JPEG will show compression artifacts or pixelation at journal print resolution, even if it looked fine on a screen.
Design for colorblind readers by default. Don't rely on red and green alone to distinguish groups; give every group a second channel — shape, pattern, or line style — that doesn't depend on color vision.
Match the journal's actual figure specs — font, resolution, dimensions — before you submit. Journals publish these; a returned-for-revision notice over an ignorable formatting spec is a purely avoidable delay.