statistical mirages · metaphor 73 of 100
Four cancer cases on one street. Three celebrity deaths in one week. A hitter who can't miss all August. Before asking what causes the cluster, ask the question almost nobody asks: what would pure chance look like? The answer is not "evenly spread." Pure chance arrives in bunches.
Our intuition for randomness is secretly an intuition for fairness. We imagine chance as a scrupulous host seating guests — everything spaced out, everyone taking turns, no street getting two tragedies before every street has had one. Independence means events ignore each other, and things that ignore each other collide constantly. Nothing steps aside; nothing waits its turn.
It is the evenly-spaced world that would scream design — even spacing takes coordination, repulsion, a rule. So the pattern-hungry mind, meeting genuine randomness, finds "clusters" everywhere and demands explanations for what needed none. The instrument below scatters misfortune by pure chance, over and over, and lets you watch your own alarm go off.
What chance looks like
When events are independent — when this house's misfortune knows nothing of that one's — the count landing in any window follows the Poisson distribution, and Poisson counts are lumpy by nature: their variance equals their mean, so wide swings around the average are the norm. Some streets get nothing. Some street gets four. No mechanism required — only the absence of one.
To get the smooth scatter your eye expects, events would have to consult each other: each new case checking where the others fell and keeping its distance. That is how trees space themselves in a forest (competing for light), how territorial birds nest, how parked cars fill a street. Spacing is what interaction looks like. During the V-1 bombardment of London, whole districts were sure the hits were targeted — their blocks kept getting struck in bunches. The actuary R. D. Clarke gridded the city into 576 squares and counted: the hits fit a Poisson distribution almost perfectly. The clumps were chance, wearing the mask of intent.
What to try
Hit scatter again and watch a different street get "the cancer cluster" every single time — always alarming, always circled, always somewhere. Then check the gold histogram: this map's tightest clump almost always sits squarely inside what chance owes you anyway.
Guess, get it wrong, guess again. Somewhere around round five, something shifts: you start looking for the clumps and voids instead of flinching at them. That shift is the recalibration — lumpiness stops reading as evidence.
At λ = 8 the year is mostly drought punctuated by eerie pairs; at λ = 96 the bunches merge into texture. Streaks feel most fateful exactly where events are rare — where the mind has the fewest examples of what chance can do.
In the timeline readout, the fixed-in-advance probability is often a few percent while somewhere in the year sits near certainty. Those numbers describe the same events. Only the question changed.
The somewhere effect
The probability that this street gets four cases this year is genuinely tiny — the alarmed neighbor computing it is not wrong. But nobody surveys one pre-registered street. The eye sweeps the whole town, every window at once, and settles on the densest patch it can find; the news sweeps every town. The probability that some clump somewhere crosses the alarm threshold is close to one, because a search over thousands of windows buys thousands of lottery tickets against a tiny per-window chance. The cluster did not find you. You found it — with an instrument exquisitely tuned to find one.
This is why the scan circle in the town map is drawn honestly: it does exactly what your attention does, hunting the tightest window after the fact, and it never comes home empty. The only fair comparison is not "how unlikely is this clump?" but "how big a clump would the best clump of a purely random town be?" — the gold distribution. Judged against the maximum, not the average, almost every headline cluster goes quiet.
When to worry
None of this means clusters are never real — cholera clustered around the Broad Street pump because it was the pump. It means the question must be asked in the right order. First, fix the window before looking: decide in advance which street, which month, which team, so the somewhere effect can't sneak into the arithmetic. Second, compare against the null maximum: simulate the innocent world many times, record its tightest clump each time, and ask whether yours beats what pure chance routinely produces — not whether it beats the average. Third, build the null from the real base rate — population, age, exposure — not from flat uniformity. Only if the clump survives all three do you call the epidemiologist. Sometimes it survives. That is precisely how you would know.
The mapping
| Mathematics | Life |
|---|---|
| uniform random events | Misfortunes that ignore each other — no contagion, no curse, no coordination. |
| the clump | The street, the week, the season that got several — chance's normal texture, read as a message. |
| the even scatter | What we wrongly expect of innocence; in fact the fingerprint of interaction and design. |
| the scan window | The mind's searchlight, sweeping every street and settling on the densest patch it can find. |
| the null max-clump | How big a bunch chance owes you anyway — the honest bar a real cluster must clear. |
| the rescatter | Why next year's alarming cluster will be somewhere else, equally alarmed, equally innocent. |
Where the metaphor tears
Contagion, a shared well, a common employer, a leaking tank — real causes make real clumps, and John Snow saved lives by taking one seriously. Poisson is the null, not the verdict: it tells you when a clump demands no explanation, never that explanations don't exist. A lens that preempts investigation has failed; this one should sharpen it, by showing exactly which clumps have already been explained by nothing.
The town map spreads risk perfectly flat, but real rates never are: people cluster, exposure clusters, age and poverty cluster. A "cancer cluster" over a dense neighborhood may be nothing; the same count over empty farmland may be everything. The honest null must model the base rate — cases per person at risk, not per square mile — or the debunking becomes its own mirage.
For the person on the street with four cases, "chance owed somebody this" is true and almost useless. Statistics explains the population of clusters, not the grief inside this one; it can retire the guilt and the hunt for a curse, but it does not lighten what happened. The somewhere effect comforts the town, not the house.