The mean describes total value divided by observation count; the median describes the middle ordered observation. Use the mean for additive workload relationships and the median for a middle-case view, while inspecting tails and missing cases. Neither summary alone describes variation or validates unusual observations.
Ask what the summary needs to support
For service duration, the mean helps connect total observed time with the number of cases. The median tells the duration at the middle of the ordered set. A manager estimating aggregate recorded effort may need the mean, while a customer asking about a middle experience may find the median useful. Name the decision rather than choosing whichever statistic gives the more favourable appearance.
Keep the underlying population consistent. Completed cases exclude work still open, which can make durations look shorter during a backlog build-up. Calendar elapsed time differs from active handling time, and business-hour duration differs from continuous hours. State the start event, end event, unit and exclusions before comparing either mean or median across periods. A carefully calculated statistic remains misleading if the case population changes silently.
Inspect the distribution and the unusual cases
Sort values or inspect a suitable distribution chart. A long right tail can pull the mean above the median, while the median may remain stable despite a small number of very delayed cases. Neither behaviour is an error in the statistic. Investigate extreme records for unit mistakes, duplicated joins or genuine exceptional circumstances before considering correction. Do not delete a long case simply to make the two summaries agree.
For an even observation count, document that the median uses the midpoint of the two central ordered values under the chosen convention. Keep missing durations out of measured-time summaries only under an explicit eligibility rule, and report how many cases lack valid endpoints. Replacing missing times with zero changes both the distribution and the apparent service experience without evidence of instantaneous resolution.
Pair a central value with tail information
A mean or median can conceal consequential delays. Add counts and a relevant tail measure, such as the number above a stated service threshold, when the decision requires it. For small populations, individual long cases may be more interpretable than a finely reported percentile. Use a precision level justified by the measurement; displaying many decimals does not create more reliable timing evidence.
NIST's measures-of-location reference provides the definitions of common central summaries. The example below uses invented durations and independent arithmetic. When tracking improvement, compare like populations and show whether the change comes from ordinary cases, extreme cases or altered mix. A lower median does not establish that every customer improved, and a higher mean does not by itself identify the operational cause.
If reporting a target, say whether it applies to every case, the mean or a chosen percentile. Meeting a three-hour median target can coexist with severe delays for some customers, so the target statement must not imply universal completion within three hours.
Fictional service durations: one long case changes the mean
A fictional scenario to explain the method, not a customer result or a market benchmark.
A fictional service completes five comparable cases in 2, 3, 3, 4 and 28 hours. Total elapsed duration is 40 hours across the cases, so the mean is 8 hours. The ordered middle observation is 3 hours, giving a median of 3. One of five cases, or 20%, exceeds a stated ten-hour threshold.
The source confirms that the 28-hour case waited overnight for a required customer response. The analyst retains it because the metric measures calendar elapsed time. A separate active-handling metric would need its own event data rather than subtracting a guessed waiting period. The report shows mean, median and threshold count to explain the different aspects of experience.
These fictional observations demonstrate distribution shape, not a population service promise. Five completed cases provide limited evidence about future customers, and open cases remain outside this deliberately stated population. The two central values answer different questions without either making the long case disappear.
TAKE THIS INTO YOUR NEXT CONVERSATION
A practical next-step checklist.
- Define duration events, unit and population.
- Inspect shape and verify unusual records.
- Report missing and still-open case coverage.
- Pair central summaries with relevant tail counts.
Common questions.
Is the median always the best typical value?
No. It describes the middle observation, while additive workload questions may require the mean and total.
Should a genuine long case be removed from the mean?
Only if it is ineligible under the stated definition. Otherwise retain it and explain its influence.
Further reading & source context.
These primary references provide background. They are not evidence for the fictional example above and do not endorse Accuralytix.
Found an error or a missing qualification? Send an editorial correction.
Check my data readiness.
Use a practical tool to turn this idea into a clearer starting point.
Check my data readiness