Histogram shapes: test bins before claiming two populations
A histogram’s apparent shape depends on its bin edges as well as the observations. Before calling a cross-company dataset bimodal, test whether the pattern persists under reasonable changes in width and starting point. Keep the same underlying values, disclose normalization and distinguish a visual description from evidence that two separate economic populations exist.
Start with the research observation
Suppose an analyst plots delivery delays across companies and sees two tall bars. It is tempting to describe two types of operator, reliable and distressed. Yet a bar is an interval chosen by the analyst, not a naturally observed company category. A gap can arise from bin placement, limited observations or rounding rather than a genuine separation in business mechanisms.
The NIST histogram reference describes a histogram as counts within classes and discusses relative-frequency and density normalizations. It supports inspecting distribution shape, but it does not make any particular binning a proof of economic groups. The example below is an original illustrative research diagnostic.
A hypothetical edge change
Assume six measured delays are 0.9, 1.0, 1.1, 1.9, 2.0 and 2.1 days. With half-day bins starting at 0.5, the counts differ from the same-width bins starting at 0.75. Use left-closed, right-open intervals so observations on an edge have one unambiguous home.
| First grid interval | Count | Shifted grid interval | Count |
|---|---|---|---|
| [0.5, 1.0) | 1 | [0.75, 1.25) | 3 |
| [1.0, 1.5) | 2 | [1.25, 1.75) | 0 |
| [1.5, 2.0) | 1 | [1.75, 2.25) | 3 |
| [2.0, 2.5) | 2 | Outside those three bins | 0 |
The shifted display looks like two clusters separated by an empty interval. The original grid spreads the same six values across four occupied bars. Neither display changed a measurement. The exercise does not prove there is one population or two; it shows that a strong verbal claim needs more than one convenient arrangement of six observations.
Preserve the normalization
Equal-width count bars describe how many observations occupy each interval. If widths differ, comparing bar heights as counts can exaggerate a broad interval. For a density histogram, height is count divided by sample size and bin width, so bar area represents probability mass. A tall narrow density bar need not contain more companies than a shorter wide one.
For example, two bins containing four of eight observations each have equal mass. If their widths are one and two, density heights are 0.5 and 0.25. That is not evidence that the first group has twice as many observations. Label count, proportion or density explicitly and keep the units of the horizontal scale visible.
Test a small set of defensible choices
Show the raw values or a compact dot plot alongside a few plausible widths and origins. Do not search dozens of settings and retain only the one that makes the preferred narrative dramatic. Save the chosen settings with the dataset snapshot so a colleague can reproduce the display without guessing how the software placed its bins.
Check measurement precision and sample size. Values rounded to whole days can create spikes and empty gaps even when the underlying delays vary smoothly. A large sample does not repair mismatched definitions, such as mixing business days with calendar days. Harmonize the measurement first and retain any unresolved differences.
Investigate groups with external evidence
If a shape remains visible across reasonable displays, ask which independent business classification might explain it. Different products, contract terms or reporting conventions can create a mixture. Verify those mechanisms rather than naming the modes after looking at which issuers support the investment thesis.
Separate exploratory visualization from a fitted mixture model or formal test, each of which adds assumptions and uncertainty. A histogram can identify a useful question without resolving it. The final memo should state the sample, units, binning sensitivity and evidence still needed to interpret the shape economically.
FAQ
Is the software’s automatic bin rule objective truth?
No. A rule can be systematic and useful while remaining one display choice. Preserve it and inspect sensitivity before attaching a strong interpretation.
Does an empty histogram bin establish a real market divide?
No. Sample size, rounding and edges can produce gaps. Investigate whether separate mechanisms are supported independently.