An axis stretched to fit one extreme outlier squashes the other ninety percent of the data flat
Aliases: long-tail axes · squashed body
What it is
When extreme outliers exist in the data (one user contributing 90% of traffic), auto-scaled axes stretch the range to cover the outlier—and the body of the data (the other 90% of users) gets squashed into a nearly indistinguishable band at the bottom of the chart. The graphic looks like "one bar," but that bar actually means "one outlier plus a thousand flattened normal values." Axis compression deletes no data, yet drives the body distribution's information content toward zero—the reader can see neither the body's shape nor its internal differences.
Why it happens
The hiding mechanism is the linearity of the position channel: position faithfully maps numeric intervals, so when the numeric range is stretched by extreme values, each unit interval receives fewer pixels—with the body's 0-100 sharing an axis with an outlier's 0-10000, the body occupies 1% of the height (say, 5 pixels), and any structure within it (multimodality, central tendency) falls below visual resolution. This is not a design mistake but a visible trade-off: keeping the outlier (honesty) and keeping the body's detail (readability) are mutually exclusive on a linear axis. Visualization's mature solutions refuse the either-or: a log axis straightens multiplicative long-tail distributions (salaries, play counts—data spanning orders of magnitude become structurally visible), though it demands readers understand log semantics and is undefined for zero and negative values; an inset magnifies the body region while the main chart keeps the full range, preserving both levels at the cost of readers handling two scales; a broken axis explicitly marks the compression point, more honest than silent truncation but adding reading cost; and decomposition (main chart without the outlier plus a note "one additional value = 9,800") is clearest for narrative charts. The choice depends on reader and task: mass-audience conclusion charts favor insets or decomposition, analysis-facing work favors log axes.
Where it stops holding
Whether compression "hides" anything depends on whether the body's internals carry decision-relevant information: if internal differences within the body are irrelevant to the decision (the reader only needs to know "one whale exists"), a flattened body loses nothing; if those internal differences are the point (which user segments contribute the remaining traffic), compression seriously damages the conclusion. The log axis's applicability also needs clarification: it works only for positive distributions spanning orders of magnitude—data concentrated within one order of magnitude are actually distorted by log axes; data with zeros or negatives require transformation (shifting, sign separation) before logging, and those transformations introduce new assumptions that themselves need stating.
Applying it
- Trigger a review when an auto-scaled chart's range spans more than two orders of magnitude: is the body squashed unreadably? If so, switch to a log axis, an inset, or decomposition.
- Any truncation or compression must be explicitly marked on the chart (break symbol, "y-axis truncated" tag, inset note)—never silent.
- When decision-relevant body structure is lost in the compressed chart, add a second chart for the body's distribution rather than cramming both levels into one.
- Verification: show the chart to a reader and ask them to describe "the shape of the main body of the data"; if they cannot (only "there's one really big value"), the hiding has happened.