Mean vs median: choose a useful center for your data

Side by side

Calculate both before choosing what to report

Suppose seven workshop attendees take 8, 10, 11, 12, 13, 14, and 44 minutes to finish an exercise. These are invented data for this guide. The total is 112 minutes, so the arithmetic mean is 112 ÷ 7 = 16 minutes. The list is already sorted, and the fourth observation is 12, making the median 12 minutes.

Both results are correct. Six people finished in 14 minutes or less, so the median is closer to the experience of most attendees in this small example. The mean also carries useful information: the 44-minute completion contributed substantially to the group's total time. Choosing the median does not make that observation disappear from the underlying data.

Values: 8, 10, 11, 12, 13, 14, 44
Count: 7
Total: 112
Mean: 112 / 7 = 16
Median: fourth ordered value = 12

Handle an even number of observations carefully

For an odd count, one observation occupies the central position. For an even count of numerical observations, the usual median is the mean of the two central values. Add another attendee who finishes in 15 minutes. The ordered list becomes 8, 10, 11, 12, 13, 14, 15, 44. The central pair is 12 and 13, so the median is 12.5 minutes. The new mean is 127 ÷ 8 = 15.875 minutes.

The median need not be a value anyone actually recorded. A median of 12.5 is a calculated midpoint, not evidence that someone finished at exactly that time. When there are repeated values, descriptions such as half are strictly below the median can also be misleading; ties may sit at the midpoint.

Separate an unusual observation from an erroneous one

Return to the original seven observations and replace 44 with 114. The total rises by 70, so the mean rises by 70 ÷ 7 = 10 minutes, from 16 to 26. The median remains 12 because the middle position has not changed. This shows the mean's sensitivity to the magnitude of an extreme observation and the median's relative resistance.

NIST's statistical handbook discusses this distinction when comparing measures of location. Resistance is useful when a distribution has a long tail, but it is not a license to ignore inconvenient records. The 114-minute value might reflect a genuine accessibility issue, an interruption, or a data-entry mistake. Investigate its meaning and apply a defensible inclusion rule before deciding whether to exclude it.

Match the measure to the decision

If you want to tell an attendee where the middle completion time fell, the median is a natural choice. If you need the total person-minutes spent on the exercise, the mean and count recover it directly. If you are reserving a room until everyone finishes, neither center is sufficient: you need information about the slower completions and the consequences of an overrun.

This last distinction prevents a common planning error. A median wait of 12 minutes does not promise that any particular person will wait 12 minutes, and a mean service time does not automatically determine staffing needs. Arrival patterns, simultaneous demand, and variability matter. Descriptive statistics summarize observed values; planning requires assumptions about what comes next.

• Typical middle observation: consider the median and show spread.
• Total workload across observations: retain the total, count, and mean.
• A deadline that most observations must meet: examine an appropriate percentile and its uncertainty.
• A visibly split population: inspect the groups instead of relying on one center.

Do not average group means without checking their sizes

Imagine a morning session with two attendees and a mean completion time of 10 minutes. An afternoon session has eight attendees and a mean of 20 minutes. Averaging 10 and 20 gives 15, but that treats the sessions as equally weighted units. The attendee-level mean is (2 × 10 + 8 × 20) ÷ 10 = 18 minutes.

Neither calculation is mysterious once you state the unit. Fifteen is the unweighted average of two session means; 18 is the mean across ten attendees. Report which question you answered. Group medians are even less informative for pooling: their weighted or unweighted average generally does not recover the median of the combined observations.

Keep missingness and spread beside the headline number

A missing measurement is not automatically zero. If someone left before finishing the workshop, coding their unobserved completion as zero would distort both summaries. Decide how incomplete observations are represented, disclose that choice, and distinguish completed cases from everyone who started. A neat statistic cannot repair a poorly defined sample.

Finally, include the number of observations, units, collection period, and a description of spread such as the range or quartiles. Two lists can share a median while having very different extremes. When feasible, show a small plot or the raw values. The arithmetic is easy to check; the harder and more valuable work is making clear what was measured and what the chosen center means.

Common questions