How Asking 1,000 People Can Tell You What Millions Think
In 1936, the Literary Digest magazine conducted one of the most ambitious surveys in American history. They mailed out ten million questionnaires to predict the outcome of the presidential election between Franklin Roosevelt and Alf Landon. About 2.4 million people responded — an enormous sample by any measure. The magazine predicted a landslide victory for Landon. Roosevelt won 46 of 48 states.
The same year, a young statistician named George Gallup predicted Roosevelt's victory using a sample of roughly 50,000 people — about 2% the size of the Literary Digest's. He didn't just predict the winner; he also predicted, to within a few points, the margin by which the Literary Digest would be wrong. The difference between the two predictions had nothing to do with sample size and everything to do with sampling method. The Literary Digest drew its respondents from telephone directories and automobile registration lists — sources that, in 1936, skewed heavily toward affluent Americans, who disproportionately supported Landon. Gallup used quota sampling, a method that deliberately constructed a sample reflecting the demographic composition of the electorate. His sample was fifty times smaller and dramatically more accurate, because it was designed to represent the population rather than merely being large.
This episode is the founding story of modern survey methodology, and the principle it established — that a small, representative sample beats a large, biased one — remains the bedrock of polling, market research, and public health surveillance today. The mechanism that ensures representativeness is random sampling: selecting respondents through a process that gives every member of the target population an equal (or at least known) probability of being included. When the selection is truly random, the sample tends to reflect the population across all dimensions — age, gender, geography, income, political leaning — without the researcher having to deliberately match each one. The randomness does the work of representation automatically, because any systematic pattern in the population has a proportional probability of appearing in the sample.
The math behind this is the central limit theorem, which states that the average of a sufficiently large random sample will be close to the average of the population, and the distribution of sample averages across many possible samples will be approximately normal (bell-shaped) regardless of the shape of the underlying population distribution. In practical terms, this means that if you randomly sample 1,000 people from a population of 300 million, the proportions in your sample will typically be within about 3 percentage points of the true population proportions. That 3-point range is the "margin of error" reported in polls, and it comes not from the size of the population but from the size of the sample. Counterintuitively, a random sample of 1,000 is roughly as accurate for a country of 300 million as it is for a city of 300,000. The population size barely matters; the sample size and the randomness of the selection are what determine accuracy.
This is why pollsters don't need to survey millions of people to make reliable predictions. A well-constructed random sample of 1,000-2,000 respondents, weighted to correct for known demographic imbalances in response rates, produces estimates that are accurate to within a few percentage points the vast majority of the time. The margin of error isn't a guess or a hedge — it's a mathematical property of random sampling, derivable from first principles, and it works every time the sampling is genuinely random.
The failures are as instructive as the successes. The 2016 U.S. presidential election is often cited as a polling failure, but the national polls were actually quite accurate — they predicted the popular vote margin within their stated margins of error. The miss was in state-level polls, several of which underestimated support for Donald Trump, particularly in the upper Midwest. Post-mortem analysis identified several causes: some voters who decided late broke heavily for Trump and weren't captured in pre-election polls; polls underrepresented voters without college degrees, who favored Trump by large margins; and some respondents may have been unwilling to disclose their preference to a pollster, a phenomenon called "shy voter" bias. In each case, the problem wasn't the math of random sampling — it was a deviation from the assumption that the sample accurately represented the voting population. The sampling frame was off, and no amount of statistical correction fully compensated.
The same principles that make polls work also apply to any situation where you want to understand a large group by examining a small one. Quality control in manufacturing uses random sampling to test products off the assembly line — you don't need to inspect every widget to know the defect rate, just a random subset. Epidemiology uses random sampling to estimate disease prevalence without testing every person in a country. Environmental science uses randomly placed sensors and sampling sites to estimate pollutant levels across a landscape. In every case, the randomness of the sample is what makes the inference valid. A non-random sample — one that's convenient, self-selected, or drawn from a biased source — can be worse than no sample at all, because it produces confident but incorrect conclusions.
Random sampling is one of those ideas that sounds too good to be true: talk to a thousand people and learn what millions think. But it works, reliably, across every domain where it's properly applied. The catch is the word "properly." The randomness has to be real, the sampling frame has to cover the population, and the response rate has to be high enough that the respondents don't differ systematically from the non-respondents. When those conditions are met, a small random sample is one of the most powerful tools humans have for understanding a world too large to observe directly.