Histogram bin width changes the conclusion

What it is

A histogram partitions a continuous axis and uses bar height or area to encode frequency or density in each interval. Histogram binning sensitivity means that the same observations can show different peaks, valleys, tails, and gaps under different widths, origins, and boundary-inclusion rules. Binning is a necessary aggregation choice, not fabricated data. The hazard is treating one parameterization as the distribution's only shape. Width is particularly consequential for claims about multimodality, skew, truncation, and heaping.

Why it happens

Each bin retains a count and discards within-bin locations. Wide bins merge neighbouring structure; narrow bins can promote finite-sample fluctuation into a jagged pattern. Even at fixed width, shifting boundaries reallocates observations near an edge and moves local peaks. This is a trade-off between smoothing bias and sampling variability. A histogram is not “kernel density estimation with zero bandwidth”: it is a piecewise-constant estimator that additionally depends on fixed boundaries. With unequal-width bins, frequency must be represented by bar area and therefore by density height, or conservation is broken.

Studying it

Rules such as Scott's and Freedman–Diaconis provide starting widths from sample size and spread under particular losses and assumptions; they do not identify one true shape for a dataset. A replication can sweep plausible widths and origins, record whether substantive claims persist, and compare with an empirical distribution, raw points, or a documented density estimate. Separate exploratory shape finding from preregistered confirmation. Tuning bins after seeing a peak and publishing only the most dramatic setting adds analyst degrees of freedom.

Where it stops holding

Centre and broad range may be stable over reasonable bins, but that stability should be checked rather than assumed. Discrete measurement, rounding, and instrument resolution can create genuine heaps that indiscriminate smoothing erases. Weighted observations require weighted counts and an explicit denominator; missing values belong in a separate state. Equal visual widths on a log axis correspond to unequal intervals in raw space, so the binning space must be named. Categorical bar charts use predefined nominal categories and do not have the same continuous-axis width choice.

Applying it

• Generate an initial width with a data-driven rule, then inspect plausible coarser, finer, and shifted boundaries; do not prescribe a fixed number of variants.
• Record width, boundaries, interval closure, count or density unit, weighting, and transformation space in the title, accessible description, or downloadable data.
• Support claims about modes, tails, or gaps with raw points, an empirical cumulative distribution, or a sensitivity view, stating which structures change with binning.
• Test shape, threshold, and anomaly tasks with realistic readers. If a critical conclusion survives only a narrow parameter region, report it as uncertain rather than an established feature.

Related

• Same group: U2.08.2 Box plots hide multimodal structure · U2.08.3 Showing the raw point distribution is more reliable
• Nearby: U2.05.2 Many points overplot; use transparency or binning · U2.09.1 Box plots compress a distribution into five statistics
• Search terms: histogram bin width · bin origin · Scott rule · Freedman-Diaconis rule · sensitivity analysis