Image by Author

The mean is the first statistic most people ever learn. Add everything up, divide by how many items you have, and you’re done. It feels reliable, objective, and precise. The problem is that the mean conceals as much as it reveals, and in several common situations it tells a story that’s almost completely backwards from the truth.

Here are five examples where the average actively misleads you.

1. The Income Illusion

Imagine ten people sitting in a bar. Nine of them earn \$40,000 a year. One earns \$4,000,000. The average income in that bar is roughly $436,000, a figure that describes nobody in the room.

This is the defining problem with income data. Because a small number of very high earners pull the mean upward, it consistently overstates what a typical person earns. The US median household income and the mean household income often differ by tens of thousands of dollars for exactly this reason. When headlines report that average incomes are rising, they may be capturing the gains of the wealthiest earners while the lived experience of most households stays flat.

The median, the middle value when all numbers are sorted, resists this distortion. For skewed data with extreme values on one end, the median is almost always the more honest summary.

2. The “Average American Family”

For decades, surveys and government reports have referenced the average American family as having somewhere around 2.5 children. That number is statistically correct and practically meaningless. No family has half a child.

The average here is arithmetically valid but describes no actual family. It emerges from combining very different households: families with zero children, two children, four children, and everything in between. When a single average is used to represent a population containing genuinely distinct groups, it can land in a gap where nobody actually lives.

This problem shows up constantly in product design, policy planning, and marketing research. Building for the average user or the average patient often means building for someone who doesn’t exist.

3. Survivorship Bias in Salary Data

A university proudly reports that graduates from its computer science program earn an average starting salary of $95,000. That sounds impressive. But consider who is included in that average.

If the calculation only captures graduates who responded to the alumni survey, and those graduates are disproportionately the ones doing well, the average will be inflated. Graduates who are unemployed, underemployed, or working outside their field are less likely to respond. The data survives a filter, and that filter systematically removes the lower end of the distribution.

Survivorship bias appears whenever you measure outcomes only for the people or things that made it through some selection process. The mean of survivors is not the mean of all participants, but it’s frequently reported as though it were. The same pattern shows up in mutual fund performance data, startup success stories, and clinical trial dropout rates.

4. Simpson’s Paradox

This one is genuinely counterintuitive. A treatment, policy, or group can perform better on average in every single subgroup and still perform worse in the overall average. That’s Simpson’s Paradox.

A well-known historical example involves UC Berkeley admissions data from the 1970s. When researchers examined overall admission rates by gender, men appeared to be admitted at a noticeably higher rate, suggesting a pattern of bias. But when they broke the data down by individual department, women were being admitted at equal or higher rates in the majority of departments. The aggregate average reversed the direction of the real trend.

How? Women were applying in greater numbers to the more competitive departments with lower acceptance rates across the board. Combining those groups without accounting for their different sizes and base rates produced a misleading overall figure.

When you aggregate across groups, the mean can point in exactly the wrong direction. The lesson isn’t that averages can’t be trusted, but that aggregating across subgroups without thinking about what those subgroups represent can produce conclusions that are the opposite of true.

5. The Average Temperature Problem

A city reports an average annual temperature of 60°F. That sounds pleasant. But imagine the city experiences brutal winters reaching 10°F and scorching summers hitting 110°F, with relatively few mild days in between. The average is 60°F, but 60°F days may be genuinely rare.

This is the problem of using a mean to summarize data that has high variability or a bimodal distribution, one with two separate peaks rather than one central cluster. The average settles in the middle of two very different states and describes neither of them accurately.

Climate data, stock returns, customer wait times, and website load speeds all run into this problem. A mean of 2.3 seconds might hide the fact that most users load the page in 0.8 seconds while a subset waits 15 seconds or more. Without knowing how spread out the data is, or how the distribution is shaped, a mean alone tells only part of the story.

What to Use Instead

None of this means the mean is useless. For symmetric distributions without extreme outliers, it remains a valuable and efficient summary. The issue is treating it as a universal default when conditions don’t support it.

Before reporting or interpreting a mean, ask a few questions. Are there extreme values pulling it in one direction? Are you averaging across subgroups that should be examined separately? Does the average correspond to a value that can actually exist in your data? Is the distribution roughly symmetric, or does it have a long tail?

Pair the mean with the median when distributions are skewed. Report the standard deviation or interquartile range alongside any average so readers understand how spread out the underlying data is. Break aggregated data into meaningful subgroups before drawing conclusions.

Statistics earns its value through accurate representation, not convenient summaries. The mean earns its place when conditions suit it, and loses credibility fast when they don’t.