If ten people are in a room and an eleventh walks in — a billionaire — the room's "average wealth" leaps to astronomical figures without anyone becoming a cent richer. It's the classic example of how a single, poorly chosen measure tells a false story about perfectly true data.
Summarizing data is unavoidable. Nobody decides by staring at ten thousand rows; they look at one number that represents them. The trouble is that every summary number — mean, median, standard deviation — carries silent assumptions about the shape of the data. Choosing badly isn't a technical detail: it's the difference between a report that clarifies and one that misleads.
This article walks through the most common summary measures, shows when each one helps, when it deceives, and how to combine them to describe your data honestly — without collapsing a table full of nuance into a single misleading number.
The mean: useful, but sensitive to extremes
The mean adds up every value and divides by the number of observations. It's the most intuitive measure and, when the data are symmetric and free of extreme values, it summarizes them very well. The average salary of a homogeneous team, the average temperature of a stable week — in those cases, the mean tells you almost everything you need to know.
Its weakness is also its definition: because it uses every value, a single extreme drags it along. A customer who spends a hundred times more than the rest, an atypical month, a data-entry error with an extra zero — any of these pulls the mean away from what is "typical". When you hear "on average", it's always worth asking: average of what, and with what tails?
The median: the middle value that resists outliers
The median is the value that sits exactly in the middle once you sort the data: half the observations fall below it, half above. It adds nothing up, so an isolated extreme doesn't move it. In that room with the billionaire, the median wealth barely changes — it still describes the typical person.
That's why the median is the right measure for skewed data: incomes, house prices, response times, order values. Whenever a minority of very high (or very low) values distorts the mean, the median tells the story of the distribution's real center.
When mean and median disagree (and what that reveals)
The relationship between the two is itself information. If the mean is clearly larger than the median, the data have a right tail — a few very high values pulling the mean up. If the mean is smaller, the tail is on the left. When the two nearly coincide, the distribution is roughly symmetric.
A simple, revealing practice: always compute both. The gap between them is a cheap alarm for detecting skew, before you even draw a chart. If mean and median are far apart, you know that reporting the mean alone will mislead whoever reads you.
The mode: the most frequent category
The mode is the value that appears most often. It makes little sense for a continuous variable, but it's the natural measure for categorical data: the most-subscribed plan, the most frequent contact reason, the best-selling product. "Our typical customer chooses the mid-tier plan" is a statement about the mode, not the mean.
It's the least used of the three in numeric reports, but the most honest when the question is "what happens most often?". Confusing mode with mean — saying "on average" when you meant "most commonly" — is a subtle error that changes the meaning of the sentence.
Standard deviation: how far the data spread from the center
Two teams can have exactly the same average sales and completely different realities: one consistent month after month, the other swinging between peaks and troughs. The mean doesn't tell them apart. The standard deviation does: it measures the typical distance of each value from the mean — that is, how much the data spread out.
A small standard deviation says the values cluster tightly around the mean; a large one, that they're scattered. In roughly normal data, the empirical rule holds that about two-thirds of observations fall within one deviation of the mean and almost all within two. But mind the "roughly normal": if the distribution is heavily skewed or has several modes, that rule stops holding and the standard deviation alone misleads just as much as the mean.
The mean alone lies: always add the spread
There's a saying that sums it up: you can drown crossing a river that is one meter deep on average. The mean, with no sense of variation, hides exactly what usually matters — the risk, the inconsistency, the extremes that make the difference in practice.
The rule is to always report center and spread together. "Delivery in 48h on average, with a standard deviation of 4h" says far more than "48h on average". The first promises consistency; the second may hide 24h and 90h deliveries canceling out in the arithmetic. A center number without a spread number is a half-truth.
Common mistakes when summarizing data
- Reporting the mean of skewed data. When there's a long tail, the median describes the typical case better.
- Ignoring the spread. Two equal means can hide opposite realities; always add a standard deviation or interquartile range.
- Comparing means of very differently sized groups without weighting by the number of observations.
- Averaging percentages or ratios that are already computed. Adding up rates and dividing by their count usually gives the wrong answer; you often need to go back to the totals.
- Applying the two-thirds rule to non-normal data. Without a roughly symmetric distribution, that intuition fails.
- Letting an outlier from a data-entry error distort everything. Investigate the extremes before including them in the summary.
Mini-case: two suppliers with the same average
A retail company was comparing two carriers to renew a contract. Both showed an average delivery time of about 48 hours, and the decision seemed to come down to price. They picked the cheaper one — and stockouts in the stores kept rising.
Going back to the data, the team looked beyond the mean. The chosen carrier had an enormous standard deviation: it delivered in 24 hours as readily as in 90, unpredictably. The other delivered almost always between 44 and 52 hours. The median and the percentiles told the same story: the equal average hid opposite levels of consistency. They started including spread in the selection criteria, delivery variability fell — and with it, the stockouts.
In practice
Never report a summary number on its own. For the center, use the median when the data are skewed and the mean when they're symmetric; always add a measure of spread, whether the standard deviation or the interquartile range; and, before trusting any summary, look at the distribution. A simple histogram reveals in seconds what three statistics alone can hide — and keeps your report from accidentally telling a story the data don't support.
