Two reports show the same ten completed tasks. One reports P90 as forty-eight minutes. The other reports thirty-two minutes.
Both calculations can follow a defined method. The records have not changed. The position used to interpolate between them has.
A percentile needs a calculation record. Keep the eligible observations, their sorted order, the requested share, and the method beside the result. A label that says “P90 duration” leaves too much for the next reader to reconstruct.
Every record below is fictional. The worksheet is a proposed reporting practice. The examples describe completed observations, with no claim about future completion times.
Separate the Number From the Question
A percentile value has the same unit as the measurement. In this example, a P90 result is a duration in minutes. A sample quantile is the value calculated from an observed sample at a selected share, such as 0.90.
A percentile rank goes in the other direction. For this report, define it as the percentage of measured observations at or below a supplied duration. State that convention explicitly because treatment of ties matters.
A target threshold is a business choice. Someone might select thirty minutes as a review threshold before examining the records. That decision does not make thirty minutes the sample's P90 under every calculation method.
Keep these questions in separate report fields:
| Question | Input | Output |
|---|---|---|
| What is P90 under the recorded method? | Observations and share 0.90 | Duration in minutes |
| What share finished at or below thirty minutes? | Observations and threshold thirty | Count divided by measured count |
| Which duration should trigger review? | An operating decision and its rationale | Chosen threshold |
The mean and median review covers central summaries. Here, the task is narrower: make a selected sample quantile reproducible.
Do not turn the percentile label into a delivery commitment. The completed records describe what happened inside their measurement boundary. A service promise needs its own decision and evidence.
Freeze the Ten Observations
Define one observation as one distinct task completed during the reporting window. Measure elapsed minutes from recorded start to recorded completion. Include waiting inside that interval. Do not mix these values with active work minutes.
For the fictional extract, use completions from 2026.10.20 00:00 UTC to before 2026.10.27 00:00 UTC. Ten eligible tasks have valid durations. No duration is missing. No record is excluded after inspection.
Keep identifiers beside the sorted values:
| Sorted position | Task | Elapsed minutes |
|---|---|---|
| 1 | Q01 | 4 |
| 2 | Q02 | 6 |
| 3 | Q03 | 6 |
| 4 | Q04 | 8 |
| 5 | Q05 | 10 |
| 6 | Q06 | 12 |
| 7 | Q07 | 15 |
| 8 | Q08 | 18 |
| 9 | Q09 | 30 |
| 10 | Q10 | 50 |
The sorted list is 4, 6, 6, 8, 10, 12, 15, 18, 30, 50. The sample size, N, is ten. The two six-minute tasks remain separate observations.
For this worksheet, equal values can use task identifier as a secondary sorting key. That makes the displayed record order stable. It does not change the duration at either tied position.
If an eligible task lacks a completion timestamp, retain it in the missing-record count. Do not assign zero minutes to make the formula run. Decide whether the incomplete evidence permits a measured-subset report. Label that subset before calculating it.
Calculate Two Specified Methods
NIST's percentile guidance describes several interpolation methods. Two are R6 and R7. They use different positions in the ordered sample.
For a requested share p between zero and one, calculate the position h:
- R6:
h = p × (N + 1). - R7:
h = 1 + p × (N - 1).
For either method, if h is at or below one, return the minimum. If h is at or above N, return the maximum. Otherwise, write h = k + d, where k is the integer part and d is the fractional part. Use one-based positions.
Calculate x[k] + d × (x[k+1] - x[k]). When d is zero, return x[k]. These formulas and endpoint rules follow the NIST definitions. They do not assert a default for your application.
For P90, set p to 0.90. R6 gives h = 0.90 × 11 = 9.9. Position nine holds thirty minutes. Position ten holds fifty.
The interpolated value is 30 + 0.9 × (50 - 30) = 48 minutes.
R7 gives h = 1 + 0.90 × 9 = 9.1. The same neighboring values now receive a different interpolation fraction.
The result is 30 + 0.1 × (50 - 30) = 32 minutes.
| Record | R6 | R7 |
|---|---|---|
| Requested share | 0.90 | 0.90 |
| Position h | 9.9 | 9.1 |
| Lower and upper positions | 9 and 10 | 9 and 10 |
| Fraction d | 0.9 | 0.1 |
| Neighboring minutes | 30 and 50 | 30 and 50 |
| Calculated P90 | 48 minutes | 32 minutes |
Neither interpolated duration appears as an actual task. Nine observations are at or below thirty-two minutes. Nine are also at or below forty-eight. The method comparison changes the reported quantile while preserving the underlying ten durations.
Check the Middle and the Ties
Use additional calculations to check that the position convention is understood. They are arithmetic checks on this example, not a separate statistical validation procedure.
At p equal to 0.50, R6 gives 0.50 × 11 = 5.5. R7 gives 1 + 0.50 × 9 = 5.5. Both interpolate between ten and twelve minutes, producing eleven.
At p equal to 0.20, R6 gives position 2.2. R7 gives position 2.8. Positions two and three both contain six minutes. Both methods therefore produce six despite their different interpolation fractions.
Preserve the ties. Replacing the list with its distinct values would remove an observation and change N. A deduplicated duration list would answer a different question from the ten-task report.
Keep full calculation precision until display. If the next extract contains fractional minutes, record the final rounding rule separately. A rounded label should not become the input to the next comparison.
When software produces the result, save its product name, version, function, arguments, and missing-value settings. Leave those fields unverified until checked. A familiar function name does not prove that two tools implement the same method.
Recalculate a Combined Population
Do not average subgroup percentiles to reconstruct the pooled percentile. The following counterexample uses the same ten records and R7 throughout.
Split the synthetic records into two invented groups. Group A contains 4, 6, 6, 8, 10. Group B contains 12, 15, 18, 30, 50. Each group has five observations.
For each group, R7 at p equal to 0.90 uses h = 1 + 0.90 × 4 = 4.6.
Group A's P90 is 8 + 0.6 × (10 - 8) = 9.2 minutes. Group B's P90 is 30 + 0.6 × (50 - 30) = 42 minutes.
The average of those two values is (9.2 + 42) / 2 = 25.6 minutes. Weighting by their equal counts produces the same average.
The pooled ten-record R7 P90 remains thirty-two minutes. An average of 25.6 does not reproduce it.
| Population | Measured tasks | R7 P90 |
|---|---|---|
| Group A | 5 | 9.2 minutes |
| Group B | 5 | 42 minutes |
| Combined observations | 10 | 32 minutes |
| Average of group P90 values | Two summaries | 25.6 minutes |
Retain the underlying observations when a combined calculation may be needed. If only subgroup summaries survive, mark the combined percentile unavailable. Do not invent a pooled value from insufficient inputs.
Copy the Percentile Record
Use this proposed worksheet for each published percentile. Keep it with the source extract so a second reader can follow the complete calculation.
| Field | Entry for this fictional report |
|---|---|
| Reader question | Describe the upper part of completed elapsed durations |
| Eligible population | Distinct tasks completed inside the stated window |
| Time window | 2026.10.20 00:00 UTC to before 2026.10.27 00:00 UTC |
| Measurement and unit | Start to completion, including waiting, in minutes |
| Eligible and measured counts | Ten eligible; ten measured; zero missing |
| Exclusions | None |
| Source version | synthetic-quantile-v1, Q01 through Q10 |
| Ordered observations | 4, 6, 6, 8, 10, 12, 15, 18, 30, 50 |
| Requested share | p = 0.90 |
| Method | R7; h = 1 + p × (N - 1); linear interpolation; endpoint clamps |
| Calculation | h = 9.1; 30 + 0.1 × 20 |
| Software and version | Hand calculation; no application default assumed |
| Value and display | 32 minutes; no display rounding needed |
| Interpretation limit | These ten completed tasks only; no future guarantee |
Put a reference to this definition in the business metric dictionary. A recurring report should point to a stable definition rather than relying on whoever last exported it.
Keep the source version distinct from the calculation version. A change to either can alter a result. Record which changed before describing a movement as an operational improvement.
Review a Disagreement Before Explaining It
Suppose a second analyst reports forty-eight minutes for this extract. Ask for the sorted values and method record. Confirm the same ten identifiers before debating the result.
If one calculation uses R6 and the other uses R7, document that difference. Choose and record a common method for a like-for-like comparison. Keep the original labeled results in the review history.
If the sorted values differ, inspect eligibility and measurement first. One report may omit waiting or include a different completion window. Changing interpolation cannot reconcile different source records.
If the fifty-minute observation is questioned, use the outlier review to inspect its evidence. Its influence on a percentile does not establish that the record is wrong. Keep it unless the documented eligibility or correction process says otherwise.
A useful report sentence is specific: “For ten completed tasks in this extract, R7 P90 elapsed duration is thirty-two minutes.” Add the method record and missing count beside it.
Keep the method with the number.
Sources
- NIST/SEMATECH e-Handbook: Percentiles. Ordered samples, interpolation, and R6 and R7 definitions. Checked 2026.10.28. All durations, subgroup calculations, worksheets, and reporting decisions are original synthetic examples.