How to Analyze Business Data When Values Are Capped, Binned, or Reported as Ranges

A value that looks numeric may not be a measurement of one exact quantity.

“$1 million+,” “$250,000–$500,000,” and a value recorded as exactly $1 million because the source caps everything above that amount each tell different stories. Treating them all as ordinary numbers creates false precision, especially in totals, averages, rankings, and models.

The useful starting point is not a cleanup rule. It is a classification question: what does this field actually reveal?

Three forms of partial information

A capped or top-coded value tells you that the true value reaches a limit, but not how far beyond it goes. If a source records every value at or above $1 million as $1 million, the record means “at least $1 million.” It does not mean the value was exactly $1 million.

A binned value places a record in a defined category, such as $1 million–$5 million or 250–999 employees. A reported range also provides lower and upper bounds, but its meaning may differ. It might reflect a survey response, rounding convention, negotiated estimate, or privacy rule rather than a designed reporting band.

Documentation matters here. A value equal to a published ceiling might be a genuine amount, a disclosure-controlled replacement, or a code produced by another rule. Do not label it censored until the data dictionary supports that conclusion.

Preserve the interval instead of inventing a number

Keep the original source expression. Then create fields that make its meaning usable:

• lower bound and upper bound
• whether each endpoint is inclusive
• whether the interval is open-ended
• censor direction, if applicable
• reporting rule or source version

For example, retain the raw value “$500K–$1M,” while storing a lower bound of 500,000 and an upper bound of 1,000,000. Store “$1M+” with a lower bound of 1,000,000 and no upper bound.

Avoid replacing either field with a midpoint during import. A midpoint may be a planning assumption later, but it is not information supplied by the source. That distinction becomes crucial for wide bins and highly skewed measures such as revenue or asset values.

Ask questions the data can answer directly

Many business decisions do not require exact values.

Suppose supplier headcount is reported as 1–49, 50–249, 250–999, and 1,000+. If enhanced review begins at 250 employees, the first two groups are clearly below the threshold and the latter two are clearly at or above it. No imputation is needed.

Likewise, a CRM with revenue bands can report the number and weighted share of accounts known to have at least $5 million in revenue when $5 million is a band boundary. It cannot classify accounts against a $3 million threshold if that threshold cuts through the $1 million–$5 million band.

This logic also helps with medians and ranks. If cumulative counts cross 50% within a $250,000–$275,000 salary band, the defensible finding is that the median falls within that band. A record can be ranked above another only when its lower bound exceeds the other record’s upper bound. Otherwise, their order remains unresolved.

Use bounds for totals and averages

When every interval has finite endpoints, arithmetic can produce an honest range rather than a fictional exact result. Add all lower bounds for a lower total and all upper bounds for an upper total. Dividing those totals by the number of records produces bounds for the average.

Open-ended upper bands change the situation. With “$10 million+,” the lower total is still known, but the data alone do not provide a finite upper total or mean. Assigning $10 million, or even $15 million, to every such record does not solve the measurement problem. It adds an assumption.

That assumption may be reasonable for a specific planning exercise. If so, present it separately: the observed lower-bound result, the scenario value used for the open band, and how the conclusion changes under conservative and aggressive cases. If a lending limit or staffing decision is unchanged across plausible scenarios, better tail data may not be necessary.

Be cautious with models and trends

Replacing capped outcomes with the cap can distort relationships in a regression, particularly when the affected records sit in the distribution’s tail. Where a modeled estimate is necessary, methods designed for censored or interval-measured outcomes can be appropriate. Interval regression, for example, can accommodate exact values, bounded intervals, and left- or right-censored observations.

That does not make it automatic. The model’s distributional assumptions, sampling design, weights, repeated records, and reporting process still matter. A tail model or a fixed multiplier above a cap is also an assumption, not recovered fact.

Trend analysis needs the same discipline. Public-use data may change top-code thresholds, bin definitions, rounding rules, or other disclosure treatments between releases. A growing pileup at a cap can reflect a changed reporting rule rather than a change in the underlying business population.

For every reporting period, retain the applicable rule and report the share of records affected by caps, ranges, or open-ended bands. Partial measurement affecting 0.5% of records presents a different risk from partial measurement affecting 20% of high-value records.

The central habit is straightforward: report what the field establishes, not what a convenient conversion suggests. When exactness is unavailable, bounds, threshold counts, unresolved ranks, and scenario estimates can still support sound decisions—as long as their limits remain visible.