the second hundred · metaphor 177
You have a little experience — a few first dates, a handful of clients, three years at a job — and from it you've formed some sense of the truth. But how sure should that little pile actually make you? You can't go back and live the years again to check.
This is the quiet problem underneath every strong opinion built on thin evidence. Your estimate might be dead-on, or it might be an accident of the particular handful you happened to draw. A different week, a different set of people, and your "obvious" conclusion could have pointed somewhere else entirely. The honest question isn't what do I believe but how far could I be wrong, given only what I've seen?
You'd think answering that requires more data — a second sample, a control group, another life to compare against. The startling trick is that it doesn't. You can size your own uncertainty using nothing but the handful already in your hand, by treating it as a tiny stand-in for the world and drawing from it, again and again. Pull yourself up by your own bootstraps.
The mathematical idea
To know how much an estimate wobbles, you'd love to redo the whole study many times — draw a hundred fresh samples from the world and watch your answer jump around. That scatter is your uncertainty. But you get one sample. The world won't deal you ninety-nine more.
The bootstrap's move is audacious and almost cheeky: if the sample is all you know of the world, then let the sample play the world. Draw a new dataset the same size as yours, but by picking from your own values with replacement — so some points show up twice, some not at all. Compute your statistic on that. Do it a thousand times. Because your sample is your best picture of the world, resampling it mimics what re-running the study would have done — and the spread of those thousand answers estimates the spread you'd have seen for real.
No formula for the standard error, no assumption that anything is bell-shaped, no calculus. The 95% interval is just the middle 95% of your thousand re-livings — the 2.5th to the 97.5th percentile. It works for the mean, the median, a ratio, a correlation, almost any statistic you can compute, including the awkward ones no textbook has a clean formula for. You buy uncertainty with computation instead of algebra.
What to notice while it runs
The top line is your handful; every value is real, sampled from a fixed but hidden, lopsided world. Each frame the panel draws a fresh resample — you can see it in the dots, some swelling because they were picked twice or three times, some fading to hollow rings because they were skipped. The statistic of each resample drops a tick into the histogram below, and after a thousand the bootstrap distribution takes shape, with your confidence interval shaded across its middle.
Now press + live one more a few times. As the handful grows, the distribution narrows and the interval tightens — more experience, less doubt — and it tightens slowly, roughly as one over the square root of what you've lived. Press a different life to redraw a handful of the same size: the interval jumps to a new place but keeps about the same width — your doubt is a property of how much you've seen, not of the particular luck you got. And switch to the middle: the median's bootstrap comes out lumpy and stepped, a reminder that some estimates are just coarser than others, no matter how you slice them.
The mapping to a human worry
Most of what we conclude about ourselves and others rests on samples too small to name: the few relationships that taught you "what you're like in love," the handful of talks that convinced you you're bad at public speaking, the three bosses who formed your theory of work. The bootstrap is a discipline for that exact situation — a way to feel the width of your own evidence before you harden a handful into a law about yourself.
Its lesson is double-edged and humane. On one side: you do not need more data to become wiser about the data you have; honestly re-shuffling your own experience already tells you how much to trust it. On the other: the width never fully closes. A small handful yields a wide interval no confidence of tone can shrink, and pretending otherwise is just a wider version of the same error. To bootstrap your doubt is to hold your conclusions exactly as tightly as your evidence permits — no tighter.
Read as life lessons
Uncertainty isn't a mood; it's a width. You can put a number on how far your one conclusion might be off, using only the evidence that produced it. Vague unease becomes an interval.
The interval shrinks like 1/√n — punishingly slow. Doubling your certainty takes four times the experience. Small samples stay humble whether you like it or not.
Redraw and the estimate moves but the width holds. Your handful was never special — it's one sample among the many the world could have handed you. The interval remembers that.
In the wild
When a treatment effect has no tidy error formula, researchers bootstrap the trial data to put honest confidence intervals around it — standard practice in modern biostatistics.
Bagging and random forests are the bootstrap wearing a different hat: resample the training data, fit many models, and average away the variance of any single fit.
Complex estimators — inequality measures, weighted survey statistics — get their uncertainty by resampling, where deriving the standard error by hand would be hopeless.
The mapping, exactly
| Mathematics | Life |
|---|---|
| the sample | Your handful of lived experience — the only evidence you actually hold. |
| the sample as world | Treating that handful as your best picture of reality, since it's the one you were given. |
| resample with replacement | Honestly re-imagining how your experience could have gone — same world, different luck of the draw. |
| the bootstrap distribution | The cloud of conclusions you might have reached, spread out where they could plausibly have fallen. |
| the 95% interval | The width of your justified doubt — how tightly the evidence lets you hold the claim. |
| shrinking like 1/√n | The slow, unfair rate at which more experience buys more certainty. |
The honest model
There's a hidden world — a right-skewed distribution, values drawn as exp(1.45 + 0.55·Z) for a standard normal Z — and your sample is a genuine draw from it. Living one more experience really samples that world again; a different life really redraws the whole handful. Nothing on screen is scripted.
Each frame the panel builds resamples the true way: for a handful of size n, it draws n values from your sample with replacement and computes the chosen statistic. It streams up to B of these into the histogram, tracks the standard deviation of the resampled statistics (the bootstrap SE), and reads the interval straight off the sorted resamples as the empirical 2.5% and 97.5% percentiles. Every readout is measured from the resamples you can watch accumulate — change the handful and it all rebuilds from scratch.
Where the metaphor tears
The bootstrap measures how much your estimate would wobble if the sample fairly represents the world. It cannot detect that the sample was biased to begin with — a handful of only your happiest relationships resampled forever still says "love is easy," with a tight interval and total false calm. It sizes sampling luck, not the sin of a crooked sample.
With three or four points, the "world" you're resampling is almost nothing — the method dutifully returns an interval, but it's built on a picture too crude to trust, and for extremes like the maximum it can fail outright. A confident number computed from near-nothing is more dangerous than an honest shrug.
The trick assumes your data points are interchangeable and independent. Real experience isn't: your relationships shaped each other, your jobs were chained by the last one's reference. When the draws are entangled by time or influence, plain resampling breaks them apart and understates the true doubt. Some lives don't shuffle.