Histograms Lie by Default: Binning Choices That Change the Story

Ten bins show a healthy bell curve. Fifty bins reveal two distinct populations with a valley between them. Same data, opposite conclusions — and both charts were drawn by defaults. A histogram is not a photograph of your distribution; it is a model of it, with the bin width as its main parameter. Every default bin count is somebody's guess about your data, and the guess is frequently wrong.

The Bin-Width Tradeoff

Too few bins oversmooth: genuine structure (bimodality, skew, gaps) melts into a reassuring mound. Too many bins undersmooth: random sampling noise poses as structure, and every spike invites a story that is not there. The classic rules of thumb — Sturges, Freedman-Diaconis, square-root — disagree with each other routinely, which tells you everything about their authority: they are starting points, not answers.

The practical method is comparative, not formulaic: draw the histogram at three widths (say 10, 30, and 100 bins). Structure that survives all three is real; structure that appears at only one width is binning artifact. If the shape changes the conclusion, the conclusion was never in the data — it was in the parameter. This takes ninety seconds and kills an entire class of phantom findings — make it a reflex before any distribution claim leaves your notebook.

Escapes from Binning Entirely

When the distribution itself is the message, consider charts with no bins at all:

• Empirical CDF: plots "share of data below x" with zero parameters and zero information loss. Steep sections are dense regions; flat sections are gaps. Harder to read at first, impossible to mislead with.
• Violin / density plots: smooth the data with a kernel instead of bins. Better than histograms for comparing groups side by side — but the bandwidth is the same tradeoff in a new costume, so vary it too.
• Strip or beeswarm plots: show every observation (up to a few thousand points). Nothing to tune, nothing hidden — the honest extreme.

Binning Hygiene Checklist

Before publishing any histogram: confirm the shape survives at least two other bin widths; check that outliers are not stretching the axis and crushing the body into three bars (clip the display range and annotate the clipped count); and make sure the underlying column passed a quality check — a "bimodal" distribution is sometimes just two different units (pounds and kilograms) sharing a column. Distributions deserve the same chart-choice discipline as trends: pick the view that shows the structure your decision depends on, and verify the structure is in the data, not the defaults.