Skip to content

Error Bar

Visual encoding — instantiates Uncertainty Explicitness

A short whisker drawn through a plotted point that shows, at a glance, how far the measurement could vary — so a data point on a chart cannot masquerade as an exact, dimensionless dot.

An Error Bar is the visual encoding of uncertainty for discrete plotted points: a whisker extending above and below a point (or beside a bar) whose length shows how far the underlying value could plausibly vary. Its distinguishing move is that it acts on the chart, at the exact place a reader's eye lands. A point on a graph is a persuasive liar — it looks like a precise, dimensionless fact — and the error bar is the mark that puts the plotted uncertainty back where the plotting removed it, per point, so the reader sees the wobble at the same instant they see the value.

Example

A materials lab measures the tensile strength of three coating formulations, three specimens each, and charts the means as three bars. Bare, the chart tells a clean story: formulation B is strongest, "the winner." Adding error bars redraws the picture honestly. B's mean is highest, but its whisker overlaps heavily with A's — the specimen-to-specimen scatter is wide enough that the ranking could flip on the next batch. C's bar is short with a tiny whisker: less strong, but far more consistent. The error bars turn a misleading "B wins" into the truthful "B and A are within noise of each other; C is the reliable one" — and they do it in the reader's peripheral vision, before any caption is read.

How it works

The bar renders a chosen numeric spread — a standard deviation, a standard error, or a confidence interval — as a length in the plot's own units, so uncertainty is read on the same axis as the value. What distinguishes it from its siblings is placement and immediacy: it is per-point, visual, and pre-attentive, so overlap between two points' bars communicates "these may not really differ" without any statistics. Its whole meaning, though, depends on one caption line stating what the bar represents — because the same-looking whisker can mean three different things.[1]

Tuning parameters

  • What the bar shows — standard deviation (spread of the data), standard error (precision of the mean), or a confidence interval. These differ by large factors; choosing the smaller one to look precise is a classic quiet distortion.
  • Bar length / confidence level — ±1 versus ±2 units, or 90% versus 95%; longer bars are more conservative and less flattering.
  • Symmetry — symmetric whiskers versus asymmetric bounds for skewed or bounded quantities, where a symmetric bar would imply impossible values.
  • Caption disclosure — whether the figure states which spread the bar encodes; omitting it makes the mark uninterpretable.
  • Clutter management — bars on every point versus representative bars; dense bars can bury the signal they were meant to reveal.

When it helps, and when it misleads

Its strength is speed: it delivers uncertainty in the same glance as the value, and overlapping bars quietly kill a lot of over-claimed differences before they are argued. Its failure modes are almost all about ambiguity and eye-tricks. An unlabelled bar is meaningless — SD, SE, and CI look identical but say different things, and the smallest can be chosen to flatter.[1] Readers also over-read the "non-overlap = significant" rule of thumb, which does not hold cleanly for standard-error bars. And a bar drawn only for the sampling scatter can lend a reassuringly precise look to data whose real error is a systematic bias the whisker never captures. The discipline that guards against this is a mandatory caption stating what the bar encodes, honest choice of spread, and not letting a tidy whisker stand in for uncertainty it does not measure.

How it implements the components

  • visual_uncertainty_encoding — it is a visual mark: uncertainty rendered as a length on the chart, read pre-attentively alongside the value.
  • confidence_representation — the bar's extent quantifies how much the plotted value could vary, at a chosen spread or level.
  • precision_discipline — it stops a plotted point from implying an exactness the data lacks, forcing the visual to admit its own noise.

It does not compute the interval it draws (uncertainty_budget) — that comes from Confidence Interval, whose output an error bar often just renders — and it does not shade a continuous forecast or model curve (stakeholder_uncertainty_translation on a band), which is Uncertainty Band.

  • Instantiates: Uncertainty Explicitness — it re-inserts uncertainty at the exact point a chart would otherwise erase it.
  • Consumes: Confidence Interval or a computed spread supplies the length the bar draws.
  • Sibling mechanisms: Confidence Interval · Uncertainty Band · Probability Estimate · Confidence Label · Forecast Range · Caveated Decision Memo · Risk Register · Assumption Register · Evidence Grade Rubric · Known Unknowns Log · Model Limitations Card

Notes

Error bar and Uncertainty Band are the two visual siblings and are easy to confuse. The bar is for discrete points — bars, dots, categories — and shows per-point spread; the band is a continuous shaded region for a line, forecast, or model curve. Reach for the bar when the data are separate measurements; reach for the band when uncertainty runs along a curve.

References

[1] On a figure, a standard-deviation bar, a standard-error bar, and a confidence-interval bar look identical but differ by large, systematic factors — so an error bar is uninterpretable unless the caption states which one it is. This is the reason figure-reporting guidance insists every error bar declare what it represents. ↩