Violin plot or box plot: which should you use?

What each shows

A box plot shows five numbers: median, the two quartiles, and the whiskers. Compact, universally understood, and it works well when you have many groups to compare.

A violin plot shows the full estimated distribution as a smoothed density, mirrored. It reveals shape, which is exactly what a box plot hides.

The case for the violin

A box plot cannot show bimodality. Two distinct subpopulations can produce a box plot identical to one unimodal group. If your data might be bimodal, a violin shows it and a box plot conceals it.

Use a violin when distribution shape is part of your finding, and when you have enough observations per group to estimate a density, which realistically means around 30 or more.

The case against

With small samples the density curve is largely an artefact of the smoothing bandwidth, not your data. A violin drawn from eight points looks authoritative and tells you almost nothing.

Violins are also wider, so they scale poorly past a handful of groups, and they are less familiar to some readers.

Usually better than both: show the points

With modest sample sizes, plot the individual data points, jittered to avoid overlap. This shows the real distribution, the real sample size, and the outliers, without any smoothing assumption.

A common and very effective combination is a box plot with the individual points overlaid. You get the summary and the raw data in one figure.

For paired data

Neither is right on its own. Use connected points so the reader can see each subject's change rather than only the group summary. See paired data.

Always state what the whiskers mean

Whisker conventions vary: 1.5 times the IQR, minimum to maximum, or a percentile range. Say which in the caption.

In Plotara

Box, violin and KDE plots, with individual points, jitter and paired connections available as overlays.