How The World Works | Distributions

A room contains ten people.

Five are twenty years old.

Five are eighty.

The average age is fifty.

Nobody in the room is fifty.

The arithmetic is correct.

The description is incomplete.

This is why distributions matter.

A single number tells you where the centre may be.

A distribution tells you what kind of world surrounds that centre.

Quick Read

A distribution describes how values are arranged across a population or dataset.

To understand a distribution, readers should look at more than the average.

Centre: Where are typical values located?
Spread: How much do values vary?
Shape: Is the distribution symmetric, skewed, flat, peaked or multimodal?
Tails: How common are extreme values?
Percentiles: Where does a value sit relative to the rest?
Subgroups: Is one apparently smooth population really several different populations mixed together?
Outliers: Are rare values errors, genuine extremes or signals of another process?

OpenStax distinguishes mean and median as measures of centre, standard deviation as a measure of spread, percentiles as measures of location, and skewness as a description of asymmetry. It also warns that in skewed distributions the mean can be pulled toward the tail and that spread should be inspected rather than inferred from the centre alone.

The central question is:

What does the shape of the whole population tell us that the average cannot?

The One-Sentence Answer

Distributions work by preserving information about how observations vary across a population, allowing us to see typical values, inequality, tails, clusters, thresholds and subgroup structure that disappear when data are compressed into one average.

The Distribution Chain

population → observed values → ordering / grouping → centre + spread + shape + tails → subgroup comparison → decision

Every compression loses information.

The mean compresses an entire dataset into one number.

Sometimes that is exactly what we need.

Sometimes it removes the pattern we were trying to understand.

The Mean Is Not “The Typical Person”

The mean is the arithmetic sum divided by the number of observations.

It is mathematically important.

But it does not automatically correspond to an actual or typical observation.

If half a class scores 40 and half scores 80, the mean is 60 even though nobody scored 60.

If a town has mostly modest incomes and a few extremely high incomes, the mean income can sit far above what most residents experience.

The phrase “the average person” should therefore trigger a question:

Average in which sense, and what does the distribution look like?

Mean, Median and Mode Are Different Questions

The mean balances all values numerically.

The median is the middle ordered value, with half the observations on each side.

The mode is the most common value or category.

In a symmetric unimodal distribution, they may be close.

In a skewed distribution, they can separate.

OpenStax notes that the mean is more affected by extreme values than the median, which is why median can be a better centre measure when strong outliers are present.

No measure is “best” without a reader job.

Spread Changes Everything

Consider two classes.

Class A scores: 58, 59, 60, 61, 62.

Class B scores: 20, 40, 60, 80, 100.

Both means are 60.

They are not educationally the same class.

Class A is tightly clustered.

Class B contains enormous variation.

A teacher designing one lesson for each class faces different problems even though the average is identical.

Spread tells us how much the population differs around its centre.

Standard Deviation: One Measure of Spread

Standard deviation summarises how dispersed values are around the mean.

A small standard deviation means values are concentrated close to the mean.

A large standard deviation means greater spread.

OpenStax notes that outliers can make standard deviation large and that in skewed distributions quartiles and other descriptions can be more informative than mean plus standard deviation alone.

The practical lesson is not “always use standard deviation.”

It is “never report centre while pretending spread does not exist.”

Range Is Simple—and Fragile

Range is maximum minus minimum.

It is easy to understand and sensitive to extremes.

If one erroneous measurement appears, the range can change dramatically.

This is why robust summaries such as the interquartile range can be useful when tails or outliers are important.

Percentiles Tell You Position, Not Percentage Correct

A percentile locates a value within an ordered distribution.

If a score is at the 90th percentile, roughly 90% of scores are at or below it.

It does not mean the student scored 90% on the examination.

OpenStax makes this distinction explicitly.

Percentiles are valuable because they tell us relative position even when raw scales differ.

But relative rank can improve even if absolute capability falls, or fall even if capability rises, depending on how the rest of the population changes.

Rank and mastery are different questions.

Skewness: The Tail Pulls the Story

A symmetric distribution has roughly similar shape on both sides of its centre.

A skewed distribution stretches farther in one direction.

OpenStax notes that in right-skewed distributions the mean is often greater than the median because large values pull the mean toward the long right tail.

This matters for income, waiting times, firm size, internet traffic and many other real-world quantities.

A few large observations can change the mean greatly while leaving the median relatively stable.

The tail is not statistical decoration.

It can dominate resource requirements and lived experience.

Latency Lives in the Tail

A website may have a fast average response time and still frustrate users because a small percentage of requests are extremely slow.

This is why the earlier article on latency emphasised high percentiles rather than average response time alone.

The same logic applies to services.

An average waiting time of two days can hide a tail of people waiting twenty.

If the tail contains high-risk cases, the average can actively mislead.

See How The World Works | Latency.

Outliers: Error, Exception or Signal?

An outlier is an unusually distant observation.

The wrong response is automatic deletion.

An outlier can be:

a measurement error;
a data-entry mistake;
a genuinely rare event;
a member of another subgroup;
evidence of a different mechanism;
the exact failure case the system needs to understand.

A bridge engineer should not dismiss extreme loads simply because they are rare if those loads determine safety. A school should not ignore a struggling student because the class average is strong. A cybersecurity team may care intensely about one rare event.

Outliers need diagnosis, not reflex.

Multimodality: One Population May Actually Be Several

Return to the room with five twenty-year-olds and five eighty-year-olds.

The average is fifty.

The distribution has two clusters.

This is a simple bimodal structure.

Multimodal distributions are warnings against inventing one “typical” case.

A school may contain two groups with very different prior preparation. A labour market may contain distinct occupation clusters. A service may have fast local users and slow remote users. A product may have casual users and professional users.

The mean can sit in a valley between groups that barely exist there.

Mixtures Create False Simplicity

Many real datasets are mixtures.

Hospital waiting times mix specialties. School scores mix prior attainment, language backgrounds and programmes. Household incomes mix employment types and family structures. Network latency mixes routes and devices.

An aggregate distribution can therefore be generated by several different processes.

Before building one explanation, ask whether one population actually exists.

Conditioning Changes the Distribution

A distribution can look one way overall and another way inside subgroups.

Average test scores may differ by school, but the distribution can change after conditioning on prior attainment. Average waiting time may look high overall because one service category handles unusually complex cases.

This does not mean subgroup analysis automatically explains causality.

It means aggregation can hide structure.

Good analysts ask which variables should define meaningful conditional distributions.

Simpson’s-Paradox Intuition

Sometimes a trend appears in several subgroups but reverses when the groups are combined—or the aggregate trend appears while subgroup trends differ.

This family of phenomena is associated with Simpson’s paradox.

The cause is not statistical magic.

Group sizes and conditioning variables change the weighting structure.

The lesson is essential:

Aggregate relationships can differ from relationships inside the populations being aggregated.

Never infer individual mechanisms from aggregate averages without checking composition.

The Ecological Fallacy

A group-level statistic is not automatically true of individuals inside the group.

A wealthy neighbourhood contains lower-income households. A high-performing school contains struggling students. A fast service has slow cases.

This is why policies targeted at “the average area” can miss actual receivers.

Distribution thinking keeps the people inside the average visible.

The Normal Distribution Is Useful, Not Universal

The bell curve is famous because many processes approximately produce it and because its mathematical properties are convenient.

But not every real-world quantity is normally distributed.

Income can be strongly right-skewed. Waiting times often have long right tails. City sizes and network activity can follow heavy-tailed patterns. Bounded test scores can pile up near maximum or minimum. Mixtures can be multimodal.

Do not force a bell curve onto a world that has another shape.

Heavy Tails Change Risk

Some distributions produce extreme events more frequently than a thin-tailed normal model would suggest.

This matters enormously for finance, insurance, network traffic, natural hazards and reliability.

If rare extremes dominate total loss, designing around the average event can be dangerous.

The system must ask:

How thick is the tail?
How much does one extreme event contribute to total risk?
Are observations independent?
Can several extremes occur together?

See How Risk Works.

The Average Can Hide Inequality

Suppose average income rises.

Who received the increase?

If all gains went to the top 1%, the average can rise while the median barely moves.

Suppose average class performance rises.

Did every student improve?

Did the strongest improve while the weakest stagnated?

Distribution is how we ask who received the improvement.

Distribution and Fairness

Two policies can produce the same average outcome and very different fairness profiles.

Policy A gives every person a modest improvement.

Policy B gives a huge improvement to a small group and harms another group, with the same average change overall.

The mean alone cannot distinguish them.

Distributional analysis is therefore central to public policy, education and service design whenever who gains and who loses matters.

Distribution and Externalities

An externality may be small on average and severe for a subgroup.

Average pollution exposure across a city can hide neighbourhoods with much higher exposure. Average congestion cost can hide commuters on specific routes. Average noise can hide people living beside the source.

Externality analysis therefore needs a distribution of receivers, not merely a total cost.

See How The World Works | Externalities.

Distribution and Common-Pool Resources

Average extraction per user can conceal concentrated use.

A few industrial users may consume most groundwater while thousands of households consume little. A small number of boats may account for a large share of catch.

Governance changes when extraction is concentrated.

The political power, monitoring needs and fairness questions are different.

See How The World Works | Common-Pool Resources.

Distribution and Opportunity Cost

The same choice has different opportunity costs for different people.

An hour waiting in a queue may be cheap for one person and extremely costly for another. A $50 fee has different consequences across an income distribution. A school timetable change affects students with different travel times and responsibilities differently.

Average cost can therefore hide distributional burden.

See How The World Works | Opportunity Cost.

Distribution and Nonlinearity

If response is nonlinear, the average input can produce the wrong predicted average output.

Suppose a threshold exists at 70.

A population average of 70 can mean everyone sits exactly at the threshold—or half are far below and half far above.

The system response can be radically different.

Nonlinearity makes distribution shape causally important.

See How The World Works | Nonlinearity.

The Jensen Intuition

When a response curve is curved rather than straight, applying the response to the average input need not equal averaging the responses across the distribution.

This is the intuition behind Jensen-type effects.

Variation itself can change the expected outcome.

Two populations with the same mean exposure but different variability can experience different average consequences when the exposure-response function is nonlinear.

Distribution is not always noise around the mean.

Sometimes distribution changes the mean outcome.

Distribution and Emergence

System-level distributions can emerge from local interactions.

City sizes, firm sizes, network degrees, waiting times and wealth can develop characteristic distribution shapes through different underlying processes.

The shape becomes a clue to mechanism.

But similar shapes can arise from different mechanisms, so distribution fitting alone does not prove causation.

See How The World Works | Emergence.

Distribution and Information Asymmetry

People can know the average while lacking the distribution.

A company advertises an average salary. Applicants do not know the range. A service reports average response time but not the slow tail. A school reports average results without subgroup distributions.

Selective summary can create information asymmetry even when the published statistic is true.

See How The World Works | Information Asymmetry.

Distribution and Defaults

A default chosen for the “average user” can be poor for much of the distribution.

If user needs are tightly clustered, one default may work well.

If needs are bimodal or highly dispersed, the same default imposes large costs on many people.

Distribution shape therefore affects whether one-size-fits-most design is defensible.

See How The World Works | Defaults.

Distribution and Common Knowledge

A public average can become common knowledge while the underlying distribution remains hidden.

Everyone knows the average score.

Everyone knows everyone knows it.

Yet nobody knows whether the group is tightly clustered or deeply divided.

Public summary can therefore coordinate beliefs around a statistic that compresses important heterogeneity.

See How The World Works | Common Knowledge.

Distribution in Education

A class average is a starting point, not a teaching plan.

Imagine three classes all averaging 65.

Class A clusters between 60 and 70.
Class B contains half near 40 and half near 90.
Class C contains most students near 70 with a small tail below 30.

The same mean demands three different interventions.

Class A may need whole-class refinement.

Class B may need differentiated pathways.

Class C may need targeted rescue for a small group while the majority continues.

Distribution is diagnosis.

Marks Are Bounded Distributions

Examination scores have a floor and ceiling.

This matters statistically.

A very easy test can create ceiling effects: strong students pile up near maximum, hiding differences in capability above what the test measures.

A very hard test can create floor effects: weaker students cluster near zero, making different underlying states look similar.

The shape of the score distribution therefore tells us something about the instrument, not only the students.

The Selection Effect

Who enters the dataset determines the distribution.

A university class is not a random sample of all young adults. Hospital patients are not a random sample of the healthy population. Customers who leave reviews are not necessarily typical customers.

Distribution shape can therefore reflect selection as much as underlying reality.

See How Sampling Works and How Research Bias Works.

The Survivorship Problem

Sometimes the missing part of the distribution is the important part.

Study successful firms and the failed firms are absent. Study university graduates and dropouts are absent. Examine aircraft returning from battle and destroyed aircraft are absent.

The observed distribution is conditional on survival.

Averages computed on survivors can therefore describe the wrong population.

Distribution and Probability

A probability distribution describes possible outcomes and their probabilities.

An empirical distribution describes observed values in data.

The two connect but are not identical.

We often use observed distributions to estimate or test probabilistic models.

This article owns the descriptive world of shape and heterogeneity; How Probability Works owns uncertainty over possible outcomes.

Distribution and Measurement

Measurement error can widen or distort a distribution.

If an instrument is noisy, observed variability combines true variation with measurement variation.

A change in testing method can change the observed distribution even if the underlying population has not changed.

Good analysis therefore asks whether distributional differences come from the world or the measuring system.

See How Measurement Works.

Distribution and Aggregation

Aggregation combines observations into larger groups.

This can reduce noise and reveal macro-patterns.

It can also erase heterogeneity.

A national average hides regions. A school average hides classes. A class average hides students. A company average hides departments. A monthly average hides peak-hour conditions.

The right aggregation level depends on the decision.

Temporal Distributions

Average demand per day can hide hourly peaks.

A power grid cares about peak load, not only average energy use. A transport system cares about rush hour. A school cares about when absenteeism clusters. A server cares about bursts.

Time distribution creates capacity requirements.

An average system can be adequately resourced and still fail at peaks.

Spatial Distributions

Average population density across a country says little about where people actually live.

Average rainfall says little about drought-prone regions.

Average access to hospitals can hide remote communities.

Geography turns distributions into maps.

Location matters because distance, infrastructure and clustering change the meaning of the same aggregate total.

Joint Distributions

Often we care about how two variables vary together.

Income and education.

Temperature and electricity demand.

Study time and score.

Traffic volume and latency.

A joint distribution preserves relationships that disappear when each variable is summarised separately.

This is where correlation and causal analysis begin—but correlation alone does not establish causation.

The Average Treatment Effect Is Still an Average

An intervention can help some people enormously, help others slightly and harm a subgroup while producing a positive average effect.

The average treatment effect is important.

But deployment may need treatment-effect heterogeneity.

Who benefits?

Who does not?

Under what conditions?

Distributional thinking turns “does it work?” into “for whom, how much, and at what part of the distribution?”

The Distribution Shift Problem

A model or policy built on one distribution can fail when the population changes.

Training data comes from one environment. Deployment occurs in another. A school intervention tested on one cohort is applied to another. A flood design uses historical rainfall while climate conditions shift.

The relationship may remain correct locally while the input distribution changes enough to alter total performance.

Systems must monitor not only outputs but whether the population itself has moved.

Distribution in AI

An AI model can perform well on average and fail systematically on a tail, subgroup or rare task.

Average benchmark score compresses a distribution of tasks.

A model may be excellent at common patterns and weak at low-frequency edge cases. Another may have lower average performance but fewer catastrophic outliers.

The correct model depends on the application’s loss distribution, not only average accuracy.

The Tail Can Own the System

In safety and reliability, rare events can determine architecture.

An elevator works normally thousands of times. Engineering still considers overload and failure. A financial institution has ordinary daily moves and rare extreme stress. A hospital handles routine cases and occasional surges.

If the consequence of the tail is large enough, the system must design for it.

“Rare” does not mean “irrelevant.”

Tail Risk and Redundancy

Redundancy often exists because the distribution contains rare failures.

If failure were impossible, backups would be waste.

If failure probability is small but consequence large, spare capacity can be rational.

Distribution therefore connects directly to resilience design.

See How Redundancy Works.

Distribution and Capacity

Capacity planning must ask what quantile of demand the system intends to serve.

Designing exactly for average demand guarantees shortage whenever demand is above average.

The right reserve depends on the cost of idle capacity versus the cost of failure at high demand.

Hospitals, grids, roads, warehouses and networks all live inside distributions of load.

See How Capacity Works.

Distribution and Second-Order Effects

An intervention can change not only the mean but the shape.

A policy may reduce extreme waiting times while leaving the median unchanged. A new teaching method may lift the weakest students most. A technology may raise top productivity while widening variance.

If evaluation reports only average change, second-order distributional effects disappear.

See How The World Works | Second-Order Effects.

The Distribution Audit

Define the population. Who or what is included?
Check selection. Who is missing and why?
Plot the data. Do not begin and end with a table of averages.
Choose a centre. Mean, median or mode—which answers the reader’s question?
Measure spread. Standard deviation, range, interquartile range or another relevant measure?
Inspect shape. Symmetric, skewed, multimodal, bounded?
Inspect tails. What happens at high and low percentiles?
Investigate outliers. Error, rare event or separate mechanism?
Check mixtures. Is one distribution hiding several groups?
Condition thoughtfully. What happens inside meaningful subgroups?
Check time. Are peaks or bursts hidden by temporal averaging?
Check space. Are local extremes hidden by geographical averaging?
Check nonlinearity. Does variance itself change expected outcomes?
Check fairness. Who gains and who loses?
Check distribution shift. Has the population changed since the model or policy was built?
Match the decision. Does the receiver care about the mean, median, tail, threshold or subgroup?

When the Distribution Lens Fails

Distribution thinking can become complexity theatre if every simple problem is buried under unnecessary statistics.

Sometimes the mean is exactly the right statistic.

If a factory needs total average material usage for procurement under stable conditions, the mean can be useful. If a distribution is tight and symmetric, mean and median may tell nearly the same story.

The point is not “averages are bad.”

The point is “know what information the average discarded before trusting it.”

The Average Is a Compression Algorithm

This is perhaps the cleanest way to remember the article.

A dataset can contain thousands of observations.

The mean compresses them into one number.

Compression is useful because human attention is scarce.

Compression is dangerous because discarded information may include the mechanism, the vulnerable subgroup or the tail that breaks the system.

The correct question is not whether to compress.

It is whether the compression preserves what the decision needs.

How Distributions Connect to the Rest of the World

Measurement: measurement creates the values whose distribution is observed.
Probability: probability distributions model uncertainty over possible outcomes.
Sampling: who enters the sample determines the observed distribution.
Bias: missing or distorted observations can change shape and centre.
Latency: averages can hide slow tails.
Capacity: peak and tail demand determine reserve needs.
Risk: extreme tails can dominate expected loss.
Nonlinearity: variation can change average outcome when response curves are curved.
Externalities: average harm can hide concentrated receivers.
Opportunity cost: burdens differ across people and therefore across the distribution.
Common-pool resources: average use can hide concentrated extraction.
Defaults: one default may fit a tight distribution and fail a multimodal one.
Information asymmetry: publishing the mean can hide the distribution.
Second-order effects: interventions can change variance and tails even when the mean changes little.

Questions a Reader Can Now Ask

What is the centre?
How wide is the spread?
Is the mean close to the median?
Is the distribution skewed?
Are there multiple peaks?
What happens in the tails?
Which percentile matters to the decision?
Are outliers errors or important cases?
Is the dataset a mixture of different groups?
What changes when I condition on a relevant subgroup?
Does the average hide inequality?
Could selection have removed part of the distribution?
Does nonlinearity make variability itself consequential?
Has the distribution shifted over time?
Does the summary statistic preserve what the receiver actually needs?

Frequently Asked Questions

What is a distribution?

It is a description of how values are arranged across a dataset or population, including their centre, spread, shape and tails.

Why can the average be misleading?

Because very different datasets can have the same mean. The mean can also be pulled by extreme values and can sit between separate clusters where few actual observations exist.

When should I use the median?

The median is often useful when data are strongly skewed or contain extreme outliers because it is less affected by their exact magnitude than the mean.

Why do tails matter?

Because rare extreme events can dominate safety, risk, capacity and user experience even while having little influence on how ordinary cases feel.

What is the most important practical habit?

Whenever an important decision is justified by an average, ask to see the distribution—or at least the median, spread, tails and relevant subgroups.

Research Basis and Further Reading

OpenStax, “Measures of the Center of the Data”, on mean, median and the influence of outliers.
OpenStax, “Measures of the Location of the Data”, on quartiles and percentiles.
OpenStax, “Skewness and the Mean, Median, and Mode”, on distribution shape and asymmetry.
OpenStax, “Measures of the Spread of the Data”, on standard deviation, variability and why graphing matters.

The Larger Idea

Averages are civilisationally useful.

Without compression, every report would be unreadable.

But compression has a moral and analytical cost.

The average patient is not in every hospital bed.

The average student is not sitting in every classroom chair.

The average commuter is not standing at every station.

The average income does not belong to most households.

The average response time does not comfort the person waiting in the tail.

So use the average.

Then unfold it.

Look at the spread.

Look at the tails.

Look for the second peak.

Look at who is missing.

Look at the subgroup that disappears when everything is combined.

The average tells you where the centre of gravity is. The distribution tells you what kind of world is being pulled around it.

How The World Works | Distributions — Why the Average Can Describe Almost Nobody