I've stopped using box plots (2021)
Whether traditional box plots are still a good way to visualize data is increasingly contested, especially when audiences are not statistically trained. Commenters argue that box plots are often misunderstood, can hide important features like multimodal distributions, and were largely a product of hand-drawn, pre-computer workflows; alternatives such as strip plots, violin plots, bee swarm/sina plots, histograms, and distribution heatmaps are seen as more intuitive and informative in many cases. Others defend box plots as compact summaries of quartiles and spread when comparing multiple groups, but even supporters concede they are rarely the clearest choice for broad or non-expert audiences.
Debate over box plots’ usefulness
- Many agree box plots are easily misread and often hide important structure (gaps, multimodality, tight clusters).
- Several argue they’re a relic of paper-era “data compression by hand”; computers remove that constraint.
- Others strongly defend them as a compact way to show location and spread (median, quartiles, outliers), especially for comparing multiple groups.
Audience understanding vs “education problem”
- A major theme: plots are communication tools; if many readers misinterpret box plots, they’re poor choices for most audiences.
- Some say this is just a training issue and reject dropping box plots because “people aren’t educated.”
- Counterpoint: some misperceptions (e.g., “longer shape = more data”) are cognitive, not easily fixed by explanation.
Alternatives: violin, strip, jitter, beeswarm, heatmaps
- Frequently suggested replacements: jittered strip plots, bee/swarm plots, sina plots, violin plots, stacked/side-by-side histograms, ECDFs, and “distribution heatmaps.”
- Several favor “box + overlaid raw points” as a pragmatic compromise.
- Critics of violin plots note sensitivity to KDE bandwidth, oversmoothing, poor comparability between groups, and visual clutter; some find them aesthetically or socially awkward.
- Raincloud/half-violin and ridge plots are mentioned as hybrids.
Statistical assumptions and misunderstandings
- Long subthread argues whether box plots “assume” Gaussian/unimodal data vs being fully nonparametric (just quartiles and whisker rules).
- There is confusion even among commenters about how quartiles and whiskers are defined, and about links to the central limit theorem.
- Some note that for multimodal or heavy‑tailed distributions, box plots can be actively misleading.
Use-case-driven defenses of box plots
- Defenders cite cases where stakeholders explicitly care about specific percentiles (e.g., 15th/85th, 25th/75th) and want simple comparisons across many groups.
- Box plots seen as best when: distributions are roughly unimodal, audience is statistically trained, and focus is on a small set of summary stats rather than full shape.
Meta: visualization goals and human factors
- Recurrent point: the priority should be clearest insight for the intended audience, not loyalty to a traditional chart type.
- Some conclude that disagreement and confusion in the thread itself bolster the case for favoring simpler, more literal distribution plots in most situations.