I sat through a product presentation once where the team shared a customer satisfaction score of 4.3 out of 5. The room nodded approvingly. When I asked to see the underlying distribution of responses, it turned out that roughly 55% of customers had given either a 5 or a 1. The 4.3 average was arithmetically correct, but it described almost nobody's actual experience. Most customers either loved the product or found it useless. The average implied that most people were comfortably satisfied — which happened to be the impression the presenting team wanted to leave. This is one of the most common ways averages mislead: not through deliberate falsification but through the mechanical property of how means respond to polarised distributions.
The Outlier Problem
The arithmetic mean is sensitive to extreme values. A small number of very large or very small values can pull the mean dramatically away from what most data points look like. The classic example: a bar has nine people earning around £25,000 a year, and one person who earns £250,000. The mean income in that bar is now about £47,500 — but nine out of ten people earn barely half that. The mean gives a misleading impression of the "typical" person in the bar.
The median — the middle value when data is sorted — is robust to this problem. It accurately reflects the experience of the person in the middle of the distribution, unaffected by the outlier. For anything with a skewed distribution (income, wealth, house prices, waiting times, insurance claim sizes), the median is almost always more informative than the mean.
When Aggregation Hides Variation
A country's average income might look respectable, but if most of the gains have accrued to the top 10%, most people's actual experience is much lower than the average suggests. A school's average exam score can improve even if scores fall for the majority of students, if a small group of high performers improves significantly. An investment with a positive average return can wipe out most investors if the distribution of returns is such that good years don't compensate for the catastrophic ones.
Our probability calculator shows the full distribution of outcomes for probability scenarios — not just the expected value (mean), but the range of likely results. Seeing the distribution is usually more informative than seeing the average alone.
Simpson's Paradox
One of the more counterintuitive ways averages mislead is Simpson's Paradox: a trend that appears in aggregate data can reverse or disappear when the data is broken into subgroups. In the 1970s, a study of UC Berkeley admissions found that female applicants were admitted at a lower overall rate than male applicants. But when examined by department, women were admitted at equal or higher rates in every individual department. The aggregate figure was misleading because women were applying disproportionately to competitive departments with lower admission rates overall.
The Base Rate Problem
Average success rates become misleading when the underlying population being averaged over is heterogeneous. "Our restaurant gets 4.2 stars on average" is less informative if 60% of reviewers are tourists visiting once and 40% are regulars. The average conceals that regulars might rate it 3.8 and tourists 4.5 — very different information with different implications for what the restaurant should do differently.
Mode vs Mean in Categorical Data
For categorical data — the most popular colour, the most common response in a survey — the mean is often meaningless. The average response on a five-point scale of 3.4 tells you less than knowing that the most common response was 2 ("disagree") and the second most common was 5 ("strongly agree"). A bimodal distribution — where there are two peaks — is very different from a distribution where most values cluster around the middle, even if both have the same mean.
What to Ask Instead
When you encounter an average, ask: what type of average is this (mean, median, mode)? What does the distribution look like — is it roughly symmetric, or is it skewed? Are there outliers? Is the dataset composed of different subgroups that should be examined separately? Is the sample size large enough to be reliable? These questions convert a single number — which may be anywhere from an accurate summary to actively misleading — into a more complete picture of what the data actually says.
Survivorship Bias: The Average Built Only from Winners
One of the most pervasive ways averages mislead in practice is through survivorship bias — the statistical effect of only counting entities that made it through a selection process, while ignoring those that didn't. The average return of actively managed investment funds, when calculated from current funds, looks considerably better than the average return of all funds that were launched in a given period. The difference is that underperforming funds tend to be closed or merged into other funds — they disappear from the dataset. The "average" is calculated only from the survivors.
The same effect inflates apparent success rates for restaurants (only the ones still open appear in reviews), businesses (bankruptcy filings don't appear in entrepreneurship success statistics), and music careers (the average income of professional musicians is calculated from musicians who are still working as such). When you hear an average success rate for any competitive field, the question to ask is whether the denominator includes everyone who tried or only those who succeeded. The answer changes the figure dramatically, and is almost never stated.