Sampling and Bias

In 1936 a magazine posted out millions of survey cards and confidently predicted the wrong winner of a US election — because the people who mailed theirs back simply weren't like the country as a whole. That same hidden slant still wrecks polls, product launches and medical studies today, and it has a name: bias.

A sample is only worth trusting if it is representative — if its mix of values genuinely mirrors the population it came from. And here is the trouble: the ways a sample can secretly fail to represent the whole are subtle, systematic, and almost invisible from the inside. This hidden tilt is called bias, and it has quietly derailed countless polls, medical studies, and multi-million-dollar product decisions — usually while everyone involved felt completely confident in their data.

The frightening part is how reasonable a biased sample can look. The numbers are real, the arithmetic is correct, the sample is often huge. Nothing screams "wrong." Yet the answer points steadily in the wrong direction, because of how the data was gathered rather than any mistake in handling it. Learning to smell bias before it bites is one of the most valuable skills in all of statistics.

The usual suspects

Bias is a systematic error — a built-in tilt in the sampling method that consistently misses the target in the same direction. It comes in a few recurring flavours:

The gold-standard defence against all of them is random sampling: give every unit in the population an equal chance of being chosen. Randomness doesn't guarantee a perfect sample, but it removes any hidden tilt — no group can systematically sneak in or slip out — so on average the sample mean \bar{x} centres on the population mean \mu. The only error left is ordinary chance, which averages out as the sample grows.

Worked example 1 — spot the bias

A researcher stands outside a busy gym and asks everyone leaving, "How many hours a week do you exercise?" She reports that the average person exercises 9 hours a week and concludes the nation is remarkably fit.

The flaw is selection bias. Her population is supposedly "the nation," but her sample can only ever contain people who go to a gym — precisely the most active slice of the country. Everyone who never exercises had zero chance of being asked. The 9-hour figure may be perfectly accurate for gym-goers and wildly wrong for everyone else. No amount of extra gym interviews fixes it; she is measuring the wrong population.

Worked example 2 — the opt-in online poll

A website posts "Vote now: is the new logo any good?" with a click-to-answer poll. After 50,000 votes, 78% say they hate it. Should the company scrap the logo?

Not on this evidence. This is textbook voluntary response bias. Nobody was selected — people chose to answer, and the ones who bother to click are overwhelmingly those with strong opinions, especially angry ones. The quietly satisfied majority never voted. Fifty thousand responses feels authoritative, but they are 50,000 self-selected responses, so the 78% measures the loudness of the annoyed, not the view of the customer base. A far smaller random sample of actual customers would be worth immeasurably more.

Worked example 3 — how randomization defends you

Suppose you genuinely want the nation's exercise habits. Instead of standing anywhere convenient, you obtain a list of the whole population and use a random process — think of drawing numbered tickets from a very large drum — to pick, say, 1,000 people, then chase down answers from all of them (not just the eager ones).

Now every person, gym fanatic or couch dweller, had the same chance of being picked. No group can systematically dominate the sample, so there is nothing to push \bar{x} consistently high or low. Any error that remains is the ordinary sampling variability we expect from measuring a part instead of the whole — and that kind of error really does shrink as you gather more data. Randomization converts a dangerous, unfixable tilt into a small, honest, well-behaved wobble.

Random vs biased, side by side

Switch between a random sample — highlighted points spread evenly across the population — and a biased one that only draws from the high end. Watch the sample mean \bar{x}: random keeps it near the true mean \mu; the biased sample drags it well off-target, and gathering more of the same biased points would never bring it back.

The most dangerous idea in sampling is "we'll just collect more data." That instinct works against ordinary sampling variability — the random wobble of \bar{x} genuinely averages out and shrinks as the sample grows. But it does nothing against bias.

Bias is a systematic error: every extra observation is drawn through the same tilted method, so it pushes in the same wrong direction as all the others. A biased sample of a million is just as biased as a biased sample of a hundred — it simply reports the wrong answer with more decimal places and more false confidence. That is exactly why bias is so dangerous and so easy to miss: the very thing that fixes random error (more data) makes a biased result look more trustworthy while leaving the tilt completely untouched. Fixing bias means fixing the method, not enlarging the sample.

During World War II, the American military studied bombers returning from missions to decide where to add protective armour. Plotting the bullet holes on the planes that came back, they saw damage concentrated on the wings and fuselage and hardly any on the engines — so the obvious call was to reinforce the wings and fuselage where the hits clustered.

The statistician Abraham Wald saw the trap. The planes he was looking at were the survivors — the ones that made it home. The absence of holes on the engines didn't mean engines rarely got hit; it meant planes hit in the engines mostly never came back to be studied. So the armour belonged precisely where the returning planes had the fewest holes. Wald had spotted a perfect example of survivorship bias: drawing a conclusion from a sample that silently excludes its most important cases. His insight reshaped how the military thought about vulnerability — and it is still the sharpest one-picture lesson in why who's missing from your sample can matter more than who's in it.

Often we try — but a truly random sample can be hard to get. You need a complete list of the population to draw from, a way to reach whoever gets picked, and cooperation from all of them (people who refuse re-introduce bias through the back door). Real surveys spend enormous effort chasing down reluctant respondents precisely because letting them quietly drop out would tilt the result. Randomness is the ideal to steer toward, and the closer you get, the more your sample earns the right to speak for the whole.

See it explained