Censored data: below a threshold does not mean zero
A value reported as “below 5” contains information that an ordinary blank does not: the value lies on a known side of a threshold. Replacing it with zero discards that information and adds an unsupported exact value. Research should preserve the inequality and propagate it into the result.
The NIST discussion of censoring uses reliability tests where some failure times remain beyond the observation endpoint. That is a right-censoring example. The disclosure-threshold example below is a separate application of the interval principle, not a claim that financial metrics follow a reliability distribution.
Distinguish censoring, missingness and truncation
Censoring means the observation exists but its exact value is restricted to a range. Left censoring can provide an upper threshold, right censoring a lower threshold, and interval censoring two boundaries. Retain each observation’s own threshold and its original reporting precision because rules can differ by entity or period.
An unavailable value with no supported threshold is ordinary missing information, not a censored zero. Truncation is different again: observations outside a rule may be absent from the dataset entirely. Keeping the censored row preserves a denominator that a truncated dataset may not reveal.
Record status, boundary, strictness, unit and source rule separately from the numeric field. “Less than 5” is not the same as “5 or less.” A confidentiality-suppressed cell may have additional rules and should not be interpreted as a detection-limit observation unless the source says so.
Hypothetical five-observation mean
Assume a fictional dataset has five nonnegative observations. Two exact values are 8 and 12. Each of the other three is explicitly reported as below 5. The target is the equally weighted mean across all five.
| Input group | Count | Sum information |
|---|---|---|
| Exact observations | 2 | 8 + 12 = 20 |
| Below-5 observations | 3 | At least 0, strictly below 15 |
| All observations | 5 | At least 20, strictly below 35 |
The mean is at least 4 and strictly below 7. Entering zeros yields 4, the lower boundary rather than a known mean. Entering half the threshold, 2.5, for each gives 5.5, but that is an imputation assumption. The interval alone does not establish a uniform distribution below the threshold.
If negative values were possible, the zero lower bound would be unjustified. Without another lower limit, this example would not provide a finite lower bound. If one observation were only known to exceed 20, a finite upper bound would likewise need additional information.
Choose an analysis consistent with the evidence
For descriptive research, bounds may be sufficient. Show the result under clearly labeled substitutions only as sensitivity scenarios. Keep exact, left-censored and right-censored counts next to the result, rather than concealing all replacements in a single numeric column.
A model-based estimator can use censoring information, but needs explicit distribution and observation-process assumptions. Censoring related to an unobserved outcome can make convenient methods inappropriate. Do not invoke a survival-model name merely because a cell has a greater-than sign; specify the target variable and why the model fits.
Preserve changes in disclosure thresholds over time. A fall in the fraction below a limit can reflect a lower reporting threshold rather than improving underlying values. Compare compatible rules or report how the change affects identifiability.
AI extraction should return the original inequality and a structured status. Test it with less-than signs, footnotes and scanned symbols: losing a single sign can turn an upper boundary into an exact measurement. Recheck affected calculations after correcting the extraction, retaining the earlier error record.
The same reported dataset can support distinct exact means. In the hypothetical example, censored values of 1, 1 and 1 produce a mean of 4.6, while 4, 4 and 4 produce 6.4. Both satisfy all three below-5 statements. This constructive comparison demonstrates why a single recovered mean is not identified without extra assumptions, even though the eligible row count is fully known.
FAQ
Is half the threshold an unbiased replacement?
Not generally. It depends on the unavailable values’ distribution and the observation process. The threshold alone cannot justify it.
Is a suppressed cell always censored?
No. Read the suppression rule. It may imply a range, confidentiality treatment or no usable numeric bound at all.