Why Outliers Can Change Averages

If Bill Gates walks into an ordinary bar, the average net worth of everyone in the room jumps into the billions, even though nobody else's bank balance moved by a cent. Statistics instructors have used that exact image for decades because it makes an abstract idea impossible to forget: the arithmetic mean is not a democratic vote among data points, it's a physical balancing act, and one heavy enough weight on the scale can drag the whole average toward it. Understanding why that happens, and what to do about it, is one of the most practically useful things anyone who reads a chart, a survey, or a salary report can learn.

What actually happens when a mean meets an outlier

The arithmetic mean is calculated by adding up every value in a dataset and dividing by how many values there are. That single design choice is the source of everything that follows: every data point contributes to the total in direct proportion to its size, no matter how far it sits from everyone else.

Picture a bar with 30 patrons whose net worth averages a modest $65,000 each. Total net worth in the room is roughly $1.95 million. Now Bill Gates, whose net worth is commonly estimated in the hundreds of billions of dollars, walks in. Adding even a conservative $130 billion to that total and dividing by 31 people produces an average net worth north of $4 billion per person. Not one of the other 30 people got richer. The number describing the "typical" person in the room is now wildly disconnected from what any typical person in the room actually has.

This is not a quirk or a flaw in the math. It is exactly what the mean is built to do: report the value each person would have if the total were split perfectly evenly. When the underlying distribution is lopsided, that hypothetical even split stops resembling reality. The technical term for this lopsidedness is skew — a distribution is skewed when its values pile up on one side and trail off in a long tail on the other, and income and wealth data are almost always right-skewed, with a long tail of very high earners stretching the mean upward.

A second way to see the same effect: imagine ten employees at a small company, nine earning $50,000 and one, the founder, earning $10 million in a good year. Summing those ten salaries and dividing by ten yields a mean salary above $1 million. That number is arithmetically correct and descriptively useless. It describes a situation that applies to exactly nobody in the company, including the founder, whose salary is not $1 million either.

Why the median doesn't budge the same way

The median is the middle value in a dataset once every value has been sorted from smallest to largest. If there's an even number of values, it's the average of the two middle ones. That's the entire definition, and it's why the median behaves so differently under stress from an extreme value.

Because the median depends only on rank, on which value occupies the middle position, it has no mechanism for reacting to how extreme the highest or lowest values are. Take the ten-salary example again: nine employees at $50,000 and one founder at $10 million. Sort those ten values and the middle falls between the fifth and sixth positions, both of which are $50,000. The median salary is $50,000, exactly what it would be if the founder earned $500,000 or $500 million instead of $10 million. The founder's exact salary is irrelevant to the median as long as it stays the highest value in the set.

Statisticians describe this property as robustness — a robust statistic is one that resists being distorted by a small number of extreme or atypical values. The median is highly robust. The mean is not robust at all, because every value, including the most extreme one, is baked directly into the calculation with full weight.

This distinction is why the U.S. Census Bureau reports median household income as its headline economic indicator rather than mean household income. Household income distributions in the United States are heavily right-skewed: a comparatively small number of very high-earning households pulls the mean well above what a household in the middle of the distribution actually brings home. Reporting the mean instead of the median would make every "typical American household" sound considerably richer than the household occupying the actual middle of the income ladder. Economists and government statistical agencies have leaned on the median specifically to avoid that distortion, a practice that goes back decades in U.S. Census income reporting.

A single ultra-high-net-worth individual entering a room of 30 people with an average net worth of $65,000 can mathematically push the group's reported average past $4 billion per person, which is exactly why economists report median income and median wealth rather than the mean whenever a distribution is this skewed.

How economists' choice of mean or median has changed public debate

The mean-versus-median distinction is not just a classroom exercise. It has shaped real public arguments about inequality and economic well-being, because the same underlying dataset can produce two very different headline numbers depending on which measure a report leads with.

During debates over wage stagnation and income inequality in the 2000s and 2010s, economists and journalists increasingly emphasized median household income and median wages specifically because mean figures were seen as painting an overly rosy picture of how a "typical" household was faring. When mean income rises because gains are concentrated among the highest earners, a headline that only reports the mean can make it look like broad prosperity is spreading when, in fact, the median household's income may be flat or barely moving. This gap between mean and median growth rates has itself become a commonly cited piece of evidence in discussions of rising inequality, since a widening gap between the two measures signals that gains are becoming more concentrated at the top of the distribution rather than being broadly shared.

The same pattern shows up in wealth reporting. Because wealth distributions are even more skewed than income distributions, typically due to the outsized concentration of financial assets, real estate, and business ownership among a small number of very high-net-worth households, mean wealth figures can run many times higher than the median. The Federal Reserve's Survey of Consumer Finances, a primary source for U.S. household wealth data, routinely publishes both figures side by side precisely because the gap between them is itself informative: a large mean-to-median gap signals a highly unequal distribution, while a small gap signals a more even one.

None of this means the mean is a "bad" statistic that should be discarded. It means the mean and median answer different questions, and picking the wrong one to report, whether by accident or to make a number look better than it is, can genuinely mislead an audience about what's actually happening to a typical household, employee, or citizen.

How statisticians formally decide if a value counts as an outlier

Not every large or small number is an outlier in the statistical sense. An outlier is a data point that lies unusually far from the rest of the dataset, far enough that it's likely generated by a different underlying process, a measurement error, or a genuinely rare event rather than ordinary variation. Deciding where "unusually far" begins requires a formal rule, not a gut feeling.

The most widely taught method is the interquartile range (IQR) rule, introduced by statistician John Tukey in his 1977 book "Exploratory Data Analysis." The interquartile range is the distance between the first quartile (the value below which 25 percent of the data falls) and the third quartile (the value below which 75 percent of the data falls) — in other words, the spread of the middle half of the data. Tukey's convention flags any value more than 1.5 times the IQR above the third quartile, or more than 1.5 times the IQR below the first quartile, as a potential outlier. Values beyond three times the IQR are sometimes labeled "extreme outliers" to distinguish them from milder, more borderline cases.

A second common approach uses the standard deviation, a measure of how spread out a dataset is around its mean. In a roughly normal (bell-shaped) distribution, about 68 percent of values fall within one standard deviation of the mean, about 95 percent fall within two standard deviations, and about 99.7 percent fall within three, a pattern known as the empirical rule. Values sitting three or more standard deviations from the mean are rare enough in a normal distribution that they typically warrant a closer look. This method works best on data that is at least approximately normally distributed; it becomes less reliable on heavily skewed data, like income, where the standard deviation itself is inflated by the same extreme values it's supposed to help identify.

Should outliers always be removed before running an analysis?

No. An outlier might be a data entry error worth correcting or removing, such as a decimal point typed in the wrong place, or it might be a real and meaningful extreme event, like a market crash, a record-breaking heat wave, or a single enormous transaction. Deleting outliers automatically, without investigating what caused them, risks erasing the most informative data point in the entire dataset rather than cleaning noise out of it.

Can a dataset have more than one kind of outlier?

Yes. Analysts commonly distinguish between univariate outliers, which are extreme on a single variable, such as one abnormally high income, and multivariate outliers, which only look unusual once several variables are considered together, such as a customer whose spending is normal but whose spending pattern relative to their stated income is not. Multivariate outliers are harder to spot because no single column of the dataset looks unusual on its own.

Fields that treat outliers as the whole point, not a nuisance

For many analysts, an outlier is something to flag, investigate, and possibly remove before running further calculations. But entire fields exist because some professionals care more about the extreme values than about the typical ones.

Insurance actuaries and risk managers are a clear example. A single catastrophic claim, a total loss from a hurricane, a major liability lawsuit, or a large fraud event, can matter more to an insurer's solvency than thousands of routine, small claims combined. Modeling the "typical" claim with a mean or median tells an actuary almost nothing useful about how much capital the company needs to survive a bad year. This is why extreme value theory, a branch of statistics specifically built to model the tail behavior of distributions, the rare, high-impact events, exists as its own specialized discipline rather than being treated as a footnote to ordinary descriptive statistics.

Climate science offers a similar case. A rolling average of daily temperatures tells you about typical conditions, but questions like "how severe could a 100-year flood be" or "what's the hottest plausible heat wave this century" require modeling the tail of the distribution directly, not smoothing it away. Financial risk management works the same way: stress tests for banks specifically simulate extreme, low-probability scenarios, because an institution that only plans for average conditions can be wiped out by a single unlikely event.

In each of these cases, discarding the outlier as noise would mean discarding the entire reason the analysis exists. The lesson generalizes beyond these specialist fields: before treating any extreme value as a nuisance to be trimmed away, it's worth asking whether the question actually being asked is about the typical case or about the tail.

Comparing measures of central tendency by how they handle extremes

Different summary statistics respond to outliers in genuinely different ways, and picking the right one depends on what question is actually being asked and how skewed the underlying data is.

Measure	Sensitivity to outliers	How it's calculated	Best used for
Mean	High	Sum of all values divided by count	Roughly symmetric data without extreme values, such as standardized test scores
Median	Low	Middle value once data is sorted	Skewed data such as income, home prices, or wait times
Mode	Low	Most frequently occurring value	Categorical data, or spotting the single most common outcome
Trimmed mean	Moderate	Mean calculated after removing a fixed top and bottom percentage of values	Data with a few known extreme values that shouldn't dominate but shouldn't be silently ignored either
Standard deviation	Very high	Square root of the average squared deviation from the mean	Best interpreted alongside a robust measure, not as a stand-alone summary of skewed data

The mean and standard deviation are the two measures most distorted by extreme values, because both incorporate every data point's exact magnitude directly into their formulas; the median and mode largely ignore magnitude altogether.

A trimmed mean deserves a closer look because it's a genuine middle path between the mean's full sensitivity and the median's total indifference to magnitude. A 10 percent trimmed mean, for instance, removes the highest 10 percent and lowest 10 percent of values before averaging what's left. This keeps most of the information in the data while preventing a small number of extreme values from dominating the result, which is why trimmed means show up in contexts like judged sporting events (dropping the highest and lowest scores before averaging) and in some economic indicators designed to be less noisy than a raw mean.

What people commonly get wrong about outliers

The most common mistake is treating "outlier" and "error" as synonyms. An extreme value is not automatically a mistake. Some of the most important discoveries and warnings in a dataset are legitimate outliers: the early cluster of unusual pneumonia cases that signals a new disease outbreak, the single security log entry that reveals a breach, the one customer transaction that exposes a pricing bug. Reflexively deleting anything that looks unusual can mean deleting the signal along with the noise.

A second common error runs in the opposite direction: assuming a reported average, without checking whether it's a mean or a median, tells you what's "normal." A news report that says average home prices rose 12 percent could be describing a modest, broad-based increase, or it could be describing a market where a handful of ultra-luxury sales dragged the mean upward while the price of a typical home barely moved. Without knowing which statistic is being used, and how skewed the underlying distribution is, a reader has no way to tell those two very different stories apart.

A third mistake is applying outlier-detection rules built for roughly symmetric, bell-shaped data to distributions that are inherently skewed. The standard deviation-based three-sigma rule assumes something close to a normal distribution; applying it uncritically to income, wealth, or web traffic data, which are naturally long-tailed, will flag far too many or far too few points as outliers, because the assumption the rule depends on doesn't hold. This is one reason the IQR method tends to be preferred for real-world, messy data: it doesn't assume symmetry the way the standard deviation approach implicitly does.

Anyone doing exploratory work on a new dataset benefits from pairing outlier detection with the kind of careful data cleaning process that investigates each flagged value before deciding whether to keep, correct, or exclude it, rather than applying a blanket rule.

The takeaway

A mean answers the question "what would each value look like if the total were divided evenly," while a median answers "what does a typical case actually look like," and those are meaningfully different questions the moment a dataset contains extreme values. Before trusting any reported average, whether it's income, home prices, response times, or customer spending, it's worth checking which statistic is behind the number and whether the underlying distribution is skewed enough that the choice matters. The Bill Gates in a bar image is memorable precisely because it makes that check impossible to forget: one extreme value, added to an otherwise ordinary room, can make every other person in it look far richer than they are, on paper, without a single dollar actually changing hands.

---

Read next:

• How Survivorship Bias Hides Failures
• What Data Cleaning Actually Involves
• Why Sample Size Changes Confidence

FAQ

Q: Why does one extreme value change a mean so much more than a median?

The mean sums every value in a dataset and divides by the count, so an extreme value contributes its full magnitude directly to that total, dragging the result toward it. The median only depends on which value sits in the middle position after sorting, so an extreme value at either end doesn't change its position at all, which is why the median stays stable even when one data point is wildly different from the rest.

Q: Why does the U.S. Census Bureau report median household income instead of mean income?

Because U.S. household income is heavily right-skewed, with a relatively small number of very high earners pulling the total upward. Reporting the mean would produce a number considerably higher than what a household in the middle of the distribution actually earns, while the median better reflects the experience of a typical household, which is why it's the Census Bureau's primary income indicator.

Q: What is the interquartile range method for detecting outliers?

Introduced by statistician John Tukey in his 1977 book "Exploratory Data Analysis," it flags any value more than 1.5 times the interquartile range above the third quartile, or below the first quartile, as a potential outlier. It remains a standard convention in exploratory data analysis because, unlike standard deviation-based rules, it doesn't assume the data is symmetrically distributed.

Q: Should extreme values always be removed from a dataset before calculating statistics?

Not automatically. An outlier might reflect a data entry error worth correcting or removing, or it might represent a genuine, important extreme event that carries real information, like a market crash or a disease outbreak. The right response depends on investigating the cause of the outlier first, not applying a blanket rule to strip out anything unusual.

Q: What is a trimmed mean and when is it useful?

A trimmed mean removes a fixed percentage of the highest and lowest values from a dataset before averaging what remains, such as dropping the top and bottom 10 percent. It offers a middle path between a mean, which is fully sensitive to extreme values, and a median, which ignores magnitude entirely, which is why it shows up in contexts like judged sporting events where a few extreme scores shouldn't dominate the result.

Q: Is the mean-to-median gap itself a useful statistic?

Yes. A large gap between a dataset's mean and its median signals that the distribution is heavily skewed, typically because a small number of extreme values are pulling the mean well above the typical case. Sources like the Federal Reserve's Survey of Consumer Finances publish both mean and median wealth figures side by side specifically because the size of that gap is informative about how concentrated a distribution is.