The mean does not tell the whole story

We say that something happens on average, and it seems the description is complete. But a mean can hide extremes, different groups, or a poorly chosen sample. The calculation need not be wrong: it may be answering a different question from the one the reader imagines. The first step is to separate the mathematical operation from the interpretation.

An example we can check

Let’s take five fictional waiting times, expressed in minutes: 2, 3, 3, 4, and 18. Their sum is 30 and the arithmetic mean is 6. The median, the central value in the ordered data, is 3. Both results are correct. However, saying that a typical wait was six minutes may give an inaccurate picture of four out of five experiences.

This example is original and educational; it does not describe any real service. OpenStax’s explanation of measures of center provides the definitions needed to check it. The mean uses the values of all observations; the median depends on their position in the order.

What happens if we change an extreme value

Let’s replace 18 with 38. Now the sum is 50 and the mean rises to 10, while the median remains 3. This is not a contest between a good measure and a bad one. We have observed that they respond differently to a change in the data set. That difference may be precisely what we need to communicate.

The mean is useful if we want to divide the total time among the number of people. The median can help if we ask for the point that leaves half of the observations on each side. If we are concerned about the worst experience, neither is enough: we should examine the distribution, its extremes, and how often they occur.

Spread is information, too

Two groups can share a mean and have very different patterns. In an imaginary example, group A contains 5, 5, and 5; group B contains 1, 5, and 9. The mean of both is 5. But the second has greater separation among observations. Summarizing them with the same figure would erase that difference.

The NIST guide to measures of scale describes ways to quantify spread. The range looks at the extremes; other measures use deviations from a center or positions within the ordered data. Each choice has different sensitivities. A dot plot may be more revealing than adding a figure without explanation.

More data do not fix every problem

Before interpreting a sample, we need to know where it came from. Asking only those who volunteered to respond can produce a composition different from that of the population we intend to describe. Increasing the number of responses obtained through that same mechanism does not guarantee that selection bias will be eliminated.

The OpenStax introduction to sampling helps distinguish population, sample, and variation. If a headline talks about all people but the study observed a specific group, the possibility of generalizing requires additional arguments. The precision of an operation does not replace the quality of the design.

Not all intervals mean the same thing

A standard deviation summarizes variation among data; a confidence interval for a mean addresses the estimation of a parameter under a procedure and certain assumptions. They are different things. NIST explains the conditions for intervals for the mean. Adding bars to a graph without indicating what they represent leaves an important ambiguity.

Simply seeing two intervals overlap or appear separate does not resolve every comparison. We need to know the design, the dependencies among observations, and the analysis used. The interpretation of a result must respect those conditions.

What to ask of a summary

A useful summary identifies the variable, its unit, the sample size, how it was selected, and which measure of center is presented. It adds some description of spread and explains how extreme values were handled. If there are relevant subgroups, showing them can help avoid misleading conclusions about the whole.

With this information, the mean takes its proper place: a specific tool within a broader description. The question is no longer whether we should believe in averages, but what question this average answers, with what data, and with what limitations.

Plot before summarizing

The NIST guide to box plots brings central position and spread together in a single representation. A box does not replace the data: it summarizes a distribution using quantiles, and the convention for the whiskers should be explained. A point marked as an outlier calls for investigating its origin; it does not authorize removing it to obtain a more convenient mean.

For a small sample like our fictional waiting times, showing the individual observations offers an advantage: anyone can recalculate the summary. In the accompanying bar chart, each bar corresponds to a wait, not a population category. We are not estimating how a service works or constructing a confidence interval.

The question what experience was typical must be distinguished from the question about the total accumulated time. The table preserves both summaries and shows what changed when the extreme value was replaced. The calculations can be reproduced without assuming a normal distribution or presenting the exercise as a study.