Whisker and outlier rules are not unique; state them with the chart

What it is

Box plots support multiple legitimate whisker and outside-value conventions. Whiskers may reach the minimum and maximum, chosen percentiles, or the most extreme observations inside fences 1.5 interquartile ranges beyond the quartiles; observations beyond those fences are then drawn separately. Even sample quartiles have multiple computational definitions. The number of points outside a box is therefore produced jointly by data and convention, not an automatic diagnosis of anomaly, error, or fraud.

Why it happens

The convention changes whisker endpoints and which observations receive separate marks. Tukey fences use the sample IQR as an adaptive scale and help explore tails, but tail weight, skew, sample size, and quantization all alter how many values cross them. Minimum-to-maximum whiskers retain the full range but do not highlight tail candidates. Percentile whiskers stabilize the displayed tail proportion yet may omit extremes from the summary. Hyndman and Fan's taxonomy of sample quantiles also explains why software can place box edges differently, especially with small samples.

Studying it

Record software, version, quantile algorithm, fence multiplier, whisker endpoints, and outside-point rendering before reproducing a plot. Rule comparisons should cover symmetric, skewed, heavy-tailed, discrete, and bounded distributions across sample sizes. Outcomes include stability of flagged candidates, misses, false interpretations, and what readers think an “outlier” means. A quality-control study additionally needs domain verification or known labels; another arbitrary statistical threshold is not ground truth.

Where it stops holding

The 1.5×IQR rule is an exploratory flag. It neither proves invalidity nor authorizes deletion. Valid extremes, entry errors, and observations from another population can all lie beyond a fence, while errors can remain inside. A declared transform or adjusted box plot may help with skew, but choosing whichever rule flags the fewest points after inspection is not defensible. Groups being compared need the same convention and quantile algorithm. If a business threshold defines an incident, distinguish that alert from a statistical outside value in both name and encoding.

Applying it

• State the quantile algorithm, whisker endpoint, and outside-point rule in the caption or accessible description—for example, “whiskers reach the most extreme observations inside 1.5×IQR fences; farther observations are shown separately.”
• Configure the plotting library explicitly and pin its version. Verify box edges, whiskers, and point counts on fixed test data across languages or tools.
• Call separate marks “observations beyond the fence” or “values to review.” Preserve their records, units, and provenance for a separate data-quality investigation.
• Expose counts or spread overlapping outside points, and make their values and rule available to keyboard and screen-reader users.
• Test whether readers distinguish a statistical flag from a domain anomaly. If every outside point is read as an error, revise the labels and explanation.

Related

• Same group: U2.09.1 Box plots compress a distribution into five statistics · U2.09.3 The box plot's strength is comparing many distributions side by side · U2.09.4 Sample size does not show in the box; annotate it separately · U2.09.5 Small-sample quartiles are unstable; box width is uninterpretable
• Nearby: U2.08.2 Box plots hide multimodal structure · U2.08.3 Showing the raw point distribution is more reliable
• Search terms: box plot whiskers · Tukey fences · 1.5 IQR · sample quantile definition