The Median Is Not the Message: How Averages Hide Reality

Mental Models
82 posts
- 1. The Median Is Not the Message: How Averages Hide Reality
- 2. Multipliers and Nullifiers: What Truly Moves Results
- 3. Scale Effects: Why What Works Small Can Fail Big
- 4. Robustness: Building a Life and Business That Can Handle Shocks
- 5. Fragility: How to Spot Systems That Break Under Stress
- + 77 more posts
Introduction
The median is not the message because a middle value is only a summary of a group. It does not describe every member of that group, reveal the shape of the data, or predict what will happen in one individual case.
The same warning applies to averages more broadly. A company reports an average salary. A hospital reports a median recovery time. An app reports average daily engagement. Each number may be accurate, yet each can conceal the reality you actually need to understand.
An average compresses many observations into one convenient figure. That compression is useful: without it, comparison would become exhausting. But compression always removes information. Two groups can share the same average while having radically different levels of inequality, risk, consistency, and opportunity.
The practical lesson is not to distrust statistics. It is to ask what a statistic preserves, what it hides, and whether it answers the question behind your decision.
What Does "The Median Is Not the Message" Mean?
The median is the value in the middle after a set of observations has been ordered. Half of the observations fall below it and half fall above it. If five project completion times are 4, 6, 7, 11, and 32 days, the median is 7 days.
That tells you something real: the typical position in this small group is much closer to 7 days than to 32. But it does not tell you that every future project will take 7 days. It does not tell you why one project took 32 days, how common such delays are over a longer period, or whether simple and complex projects have been mixed together.
The median describes a location in a distribution. It is not the distribution itself.
This distinction matters whenever people turn a group-level statistic into an individual prediction. If the median customer spends $60, a particular customer can still spend $5 or $500. If the median employee stays four years, a new hire is not destined to leave on their fourth anniversary. If the median time for an outcome is eight months, that number does not function as a countdown for one person.
Group summaries describe patterns. Individual outcomes depend on a combination of the pattern, the person's circumstances, relevant subgroups, changing conditions, and uncertainty.
Mean, Median, and Mode Answer Different Questions
People often use "average" to mean the arithmetic mean, but there are several measures of central tendency. Choosing among them is a judgment, not a cosmetic formatting decision.
The Mean
The mean is the sum of all values divided by the number of values. For the project times of 4, 6, 7, 11, and 32 days, the mean is 12 days.
The mean uses every observation, which makes it valuable for totals, budgets, and many forms of analysis. It is also sensitive to extreme values. The single 32-day project pulls the mean well above the time taken by most projects.
Use the mean when the total amount matters or when the distribution is reasonably balanced. For example, average electricity use can help estimate the total capacity a group of households requires.
The Median
The median is the middle ordered value, so unusually high or low observations have less influence on it. In the project example, the median remains 7 days even though one project takes 32.
This often makes the median more representative for skewed data such as income, home prices, or delivery times. But its resistance to outliers is also a limitation. A severe tail can create real costs even when it barely moves the median.
Use the median when you want to locate the middle observation and extreme values would make the mean look unlike most cases.
The Mode
The mode is the most frequent value or category. It can be useful even when a mean makes no sense. A shoe retailer may care more about the most commonly purchased size than the arithmetic average of all shoe sizes.
Use the mode when frequency is the important feature, especially for categories or repeated discrete values.
None of these measures is universally honest or dishonest. Each answers a different question. Trouble begins when a convenient measure is presented as though it were the whole story.
How Averages Hide Reality
An average hides reality in predictable ways. Learning to recognize them makes data easier to question without becoming cynical about it.
Averages Hide Spread
Consider two small teams with the same mean performance score of 80.
- Team A scores 78, 79, 80, 81, and 82.
- Team B scores 40, 60, 80, 100, and 120.
The central value is identical, but the teams are not alike. Team A is consistent. Team B has much greater variation. If you are assigning a routine process, consistency may matter. If you are looking for exceptional upside and can tolerate uneven results, the second distribution may be more interesting.
The average cannot tell you how tightly values cluster around it. You need measures or views of spread: the range, percentiles, interquartile range, standard deviation, or simply a plot of the observations.
Averages Hide Shape
A distribution can be symmetric, skewed, flat, clustered, or split into several peaks. The same mean or median can sit inside any of those shapes.
Imagine a product with an average rating of three stars. One possibility is that almost everyone considers it mediocre. Another is that half the customers love it and half hate it. Those realities call for different responses. The first suggests a generally underwhelming product. The second may indicate a sharp difference between customer segments, use cases, or expectations.
A single number removes the shape that would reveal the disagreement.
Averages Hide Tails
The tail of a distribution contains uncommon but extreme outcomes. Those outcomes can dominate a decision even when the middle looks safe.
Suppose a delivery service has a median delivery time of two days. Most customers may receive packages promptly, while 10 percent wait more than two weeks. If you are shipping an ordinary household purchase, the median may be enough. If you are shipping a replacement part that can stop a factory, the bad tail matters far more than the typical case.
This is a version of asymmetric risk: a rare delay can impose a cost much larger than the benefit of normal speed. The median remains true, but it is not the message your decision needs.
Averages Hide Subgroups
Mixing unlike groups can produce a summary that represents almost nobody.
Suppose a company reports an average commute of 35 minutes. Remote employees have almost no commute, office employees in one city average 25 minutes, and employees in another city average 70 minutes. The company-wide number is mathematically correct, but it gives a poor picture of any group's experience.
Segmentation can reveal the structure the combined average conceals. Useful segments might include location, age, product plan, customer type, project complexity, acquisition channel, or time period. The right segment is the one connected to the causal or practical question, not whichever division happens to make a chart look impressive.
Averages Hide Change Over Time
An annual average can conceal a trend, a sudden break, or a seasonal cycle.
A business might average 1,000 weekly orders across a year. That could mean stable demand near 1,000 every week. It could also mean orders fell steadily from 1,500 to 500. The first pattern supports predictable planning. The second signals deterioration.
Whenever sequence matters, aggregate statistics should be paired with a time series. Otherwise, improvement and decline can cancel each other out and leave a reassuring but stale center.
Averages Hide Individual Uncertainty
Even a well-chosen average does not erase variation around it. It tells you where a group is centered, not where the next observation must land.
This is especially important when the outcome is personal. A population statistic is a starting point for reasoning, not a verdict. The relevant base rate should influence your expectations, but so should information that genuinely distinguishes the case in front of you.
The balance is subtle. Ignoring the group pattern creates overconfidence in a special story. Treating the group pattern as destiny ignores meaningful individual evidence. Good judgment uses both.
A Concrete Example: The Average Salary
Imagine a ten-person company with annual salaries of:
$40,000, $42,000, $44,000, $46,000, $48,000, $50,000, $52,000, $54,000, $80,000, and $544,000.
The mean salary is $100,000. The median is $49,000.
If the company advertises an "average employee salary of $100,000," the claim is technically correct but likely to create the wrong impression. One unusually high salary pulls the mean upward. The median better represents the middle employee.
But the median is not the complete message either. It does not show that eight people earn between $40,000 and $54,000, one earns $80,000, and one earns $544,000. It says nothing about roles, hours, tenure, equity, location, or whether compensation is fair for the work.
The right analysis depends on the question:
- A candidate estimating a likely offer should compare salaries for the relevant role and level.
- A finance team calculating total payroll must use the actual sum, for which the mean remains informative.
- A leader assessing pay inequality should inspect the entire distribution and compensation ratios.
- A policymaker studying living standards may want median income, household composition, and local costs.
There is no context-free "best average." There are statistics that fit a question and statistics that do not.
When an Average Is Still Useful
The warning that averages hide reality can itself be taken too far. Summary statistics are indispensable. They let us compare large groups, detect changes, establish base rates, plan capacity, and communicate without listing every observation.
An average is often useful when:
- the distribution is reasonably stable and well understood;
- variation is small relative to the decision;
- the total quantity matters;
- you are comparing like with like;
- the statistic is paired with an appropriate measure of spread;
- the cost of unusual outcomes is low; or
- you need a starting point before gathering more specific information.
If a cafe sells an average of 200 cups of coffee each weekday with little variation, that number is useful for staffing and inventory. Examining every transaction individually would add effort without changing the decision much.
The goal is proportionality. Ask for more detail when hidden variation could change what you do. Accept a useful summary when it cannot.
Common Mistakes When Reading Averages
Treating "Average" as a Complete Description
The word should trigger a follow-up question: which average? A mean, median, and mode can lead to different impressions. If the measure is not named, you do not yet know what has been summarized.
Assuming the Average Is Typical
In a skewed distribution, the mean may describe no common case. The average customer, household, or project can be a mathematical construction rather than an actual member of the group.
Removing Outliers Automatically
Some outliers are data errors. Others are the most important events in the dataset. A fraudulent transaction, catastrophic machine failure, or viral customer may be unusual without being irrelevant.
Investigate why an observation is extreme before deleting it. The tail may reveal a broken process, a new opportunity, or the risk that controls the entire decision.
Segmenting Until the Result Looks Good
Breaking data into meaningful groups can reveal reality. Breaking it repeatedly until one group supports a preferred story is cherry-picking. Segments should be selected because they connect to a plausible mechanism or decision.
Confusing Description With Explanation
An average describes what happened in a group. It does not explain why. If customer retention averages 70 percent, the number alone does not identify the causes of churn. Moving from description to explanation requires hypotheses, comparisons, and often experiments.
Ignoring the Reference Class
An average for the wrong group can be worse than no average. The completion rate for all software projects may tell you little about a small internal tool with an experienced team. Narrow the reference class when the distinction is relevant, but do not narrow it merely to make your preferred outcome seem more likely.
How to Look Beyond the Average
Use a repeatable process whenever a central number affects an important choice.
1. Name the Decision
Begin with what you are trying to decide. Are you estimating a total, describing a typical case, preparing for downside, comparing groups, or predicting one outcome? The question determines which statistics matter.
2. Identify the Measure
Ask whether the number is a mean, median, mode, rate, or something else. Learn how it was calculated, over what period, and from which observations. Labels such as "average return" or "average response time" are incomplete without that context.
3. Inspect the Distribution
Look at a histogram, scatter plot, ordered list, or percentile table. At minimum, ask for the range and a measure of spread. Notice skew, clusters, gaps, and extreme values.
Useful questions include:
- How far apart are ordinary outcomes?
- What happens at the 10th and 90th percentiles?
- Is there one peak or more than one?
- Are extreme observations errors, rare realities, or a separate group?
4. Find Meaningful Subgroups
Separate observations when there is a sound reason to expect different behavior. Averages by customer plan may reveal that a product works well for professionals but poorly for beginners. Project estimates by complexity may be more useful than one company-wide completion time.
5. Examine the Tail That Can Hurt You
Do not focus only on the center when the downside is asymmetric. Ask how bad an unusual result can become, how often it occurs, and whether you can survive it. This is where margin of safety complements the average: planning for error matters when the cost of being wrong is high.
6. Combine the Base Rate With Case-Specific Evidence
Start with the relevant group pattern, then update it using evidence that truly applies to the current case. This avoids both base-rate neglect and the mistake of treating a population summary as an individual fate. Bayesian thinking provides a useful framework for making that update.
7. Communicate the Number With Its Limits
Good reporting does not bury every reader in statistical detail. It pairs the headline figure with the one or two facts most likely to change its interpretation.
For example: "Median delivery time was two days, but 10 percent of deliveries took longer than eight days." That sentence preserves simplicity while revealing the relevant tail.
A Practical Checklist
Before acting on an average, ask:
- Which kind of average is this?
- What observations and time period does it include?
- How much variation surrounds the center?
- Is the distribution skewed, clustered, or split into subgroups?
- Could the tail dominate the consequences of my decision?
- Does this statistic describe the reference class I actually care about?
- Am I using a group pattern as though it were an individual prediction?
- What additional number would most change my interpretation?
You rarely need every possible statistic. You need enough of the distribution to see whether the central number supports the decision you are making.
Final Thoughts
The median is not the message. Neither is the mean, the mode, or any other single summary. Averages are maps: valuable because they simplify, dangerous when their missing details contain the terrain that matters.
Use an average to orient yourself, then look deeper when variation, tails, subgroups, trends, or individual circumstances can change the choice. The habit is simple: identify the measure, inspect what surrounds it, and match the statistic to the decision.
If you want a deeper framework for using mental models in everyday decisions, 100 Mental Models expands on these ideas in a broader and more practical way.
The point is not to demand perfect information. It is to stop mistaking one useful number for the whole reality.
Key Takeaways
- An average summarizes a dataset, but it does not show its spread, shape, subgroups, or the range of possible individual outcomes.
- The mean, median, and mode answer different questions, so the right measure depends on the decision you need to make.
- Better decisions come from inspecting the distribution, relevant segments, and uncertainty instead of relying on one central number.
Quick Q&A
What does the median is not the message mean?
It means the middle value of a group is a useful summary, but it does not determine the experience or outcome of any particular member.
How can you avoid being misled by averages?
Ask which average is being used, inspect the range and distribution, separate meaningful subgroups, and match the statistic to your decision.
Part of 82 in
Mental Models