The misconception
People often use average as if it meant the observation in the middle of an ordinary experience. In statistics, average usually means the arithmetic mean, and the arithmetic mean does not promise to be ordinary. It is a balance point. Every value contributes according to its size, so one sufficiently large value can move that balance point far away from most observations.
Consider ten delivery times measured in minutes: 8, 9, 9, 10, 10, 10, 11, 11, 12, and 90. Nine deliveries form a compact group. One delivery took much longer because our fictional route generator added a temporary road closure. The mean is 18 minutes, even though no delivery took anything close to 18 minutes. The median is 10 minutes. Neither answer is a mistake. They answer different questions.
The mean answers, “What equal value would preserve the total time?” The median answers, “What value divides the ordered observations into two equal halves?” Calling either one the true middle hides that distinction.
A synthetic experiment
The linked experiment starts with nine fixed values between 8 and 12 and one adjustable final observation. Move that observation from 10 to 30, 60, and 90. The chart marks the moving point, mean, and median; the three headline metrics update the mean, median, and 10% trimmed mean after every movement. The values table also reports quartiles, the IQR outlier count, and the upper fence.
At 10 minutes, all three summaries are nearly identical. As the adjustable point moves right, the mean follows it continuously. The median barely moves because the adjustable point remains at the end of the ordered list. The trimmed mean initially follows the change, then stops using the most extreme point after trimming one observation from each tail.
Every number in this experiment is synthetic. It was chosen to expose the behavior of the summaries, not to describe any delivery company, neighborhood, or transport system. That matters because the same arithmetic can be correct while the story attached to it is unsupported.
Use the slider’s full implemented range from 8 to 100. Near the fixed group, all three centers remain close. As the final value passes the displayed upper fence, the IQR count identifies it as unusual while the mean continues moving. That flag does not remove the point or decide that it is erroneous. Only one observation is adjustable in this lab, which deliberately isolates its influence on the three summaries.
Derivation and reasoning
For observations x1, x2, ..., xn, the arithmetic mean is:
mean = (x1 + x2 + ... + xn) / n
If one observation increases by d, the mean increases by exactly d / n. In a sample of ten, changing the last observation from 10 to 90 adds 80 minutes to the total and therefore adds 8 minutes to the mean. This influence is simple, predictable, and sometimes desirable. If the goal is to budget total driver time, the 90-minute event belongs in the total.
To obtain the median, sort the observations. For an odd sample size, select the central observation. For an even sample size, average the two central observations. The actual distance from the center to the extremes does not enter that calculation. Changing the largest value from 12 to 90 preserves the ranks of the two middle values, so the median is unchanged.
There is also a useful optimization interpretation. The mean is the value that minimizes the sum of squared deviations. Squaring gives large deviations extra weight. The median minimizes the sum of absolute deviations, which grows only in direct proportion to distance. This helps explain why the mean is natural for least-squares models and why the median is robust in heavy-tailed settings.
A trimmed mean deliberately removes an equal proportion from both tails before averaging. It can balance some efficiency of the mean with some resistance to extremes, but its trimming proportion is a modeling choice that must be disclosed.
Assumptions
Before comparing centers, make sure the values are genuinely comparable. Mixing minutes and hours, combining different definitions of completion, or silently treating missing deliveries as zero makes every summary misleading. The observations also need a meaningful numeric scale. A mean of arbitrary category codes has no useful interpretation.
The shape of the distribution matters. For roughly symmetric, light-tailed data, the mean and median are often close, and the mean uses the available magnitude information efficiently. For skewed or heavy-tailed data, the difference between them is evidence worth explaining rather than a reason to hide one result.
Finally, a descriptive center does not establish a forecast. The mean of ten synthetic observations is not automatically the expected time for a future observation unless the data-generating process is stable and the sample is relevant.
Failure cases
Do not delete the 90-minute observation merely because it is inconvenient. First ask whether it is a recording error, a valid rare event, or evidence that the process has more than one operating state. A valid extreme may determine staffing, service guarantees, or risk even when it is not typical.
Do not report only the median when totals matter. A warehouse cannot replace ten observed work periods with ten copies of the median and expect the same labor cost. Conversely, do not report only the mean when the reader wants a typical personal experience. Report the distribution, sample size, and spread whenever the decision deserves them.
Multimodal data create another failure case. If half the synthetic deliveries take 10 minutes and half take 50, a center near 30 describes neither group. Segmenting by a justified operational factor or showing the full distribution may be more honest than searching for a better single number.
Source notes
The NIST/SEMATECH e-Handbook distinguishes mean, median, mode, and robust alternatives, and discusses how skewness and heavy tails affect their usefulness. OpenStax separately develops the mean and median as different measures of center. This guide uses that shared framework but supplies its own synthetic delivery example and wording. Neither source establishes a universal winner; the estimator must be connected to the distribution and the purpose of the analysis.
Keep these distinctions
- The mean uses every magnitude, while the median mainly uses order.
- An extreme value may be an error, a rare event, or the most important observation in the data.
- Choose a measure of center for the question and distribution, not by habit.
Sources checked
These sources support the statistical definitions and boundaries. The scenario, prose, calculations, and figure on this page are original to Count & Chance.
- 1.3.5.1. Measures of Location, National Institute of Standards and Technology. Checked 2026-08-19.
- 2.5 Measures of the Center of the Data, OpenStax, Rice University. Checked 2026-08-19.