How Political Polls Work: Margins, Weighting and Why They Miss

A Thousand Strangers Stand In for a Whole Country

Every election season produces the same small miracle of arithmetic: a polling firm calls roughly a thousand people, asks a handful of questions, and announces what a nation of tens or hundreds of millions intends to do — usually within a few points of the truth. Understanding how political polls work turns that miracle back into a method, and the method has four moving parts: building a sample that resembles the electorate, weighting the answers when reality refuses to match the plan, modelling who will actually vote, and reporting uncertainty honestly — the step where headlines most often betray the science. This guide walks through each part in plain language: how samples are drawn in an era when almost nobody answers unknown numbers, what weighting actually adjusts (and the quiet scandal of education weighting that explained a famous miss), what the margin of error does and does not cover, and the real reasons polls fail — which are usually about late movement and hidden voters, not incompetence. It pairs with our explainer on how international summits work — together they cover the two rituals, domestic and diplomatic, through which modern politics announces what it is about to do, and why both should be read with the same informed scepticism.

Sampling: Finding a Miniature Electorate

A poll’s raw material is a sample — a group small enough to interview and structured to mirror the voting population in miniature. The mathematics is startlingly forgiving: a random sample of about 1,000 voters represents a country of 100 million almost as precisely as it represents a city of 100,000, because precision depends on the sample’s size, not its share of the population (the same reason a cook tastes one spoonful of soup, not one percent of the pot). The craft is achieving anything close to random in practice. Traditional telephone polls dialled random numbers across landlines and mobiles; response rates have collapsed from roughly one in three in the 1990s to well under one in ten today, so modern pollsters blend methods — phone calls, online panels recruited to quota targets, text-to-web invitations, and address-based mail. Each method reaches different people: online panels skew younger and more educated, phone samples older, mail slowest. None of this is fatal — a biased sample can still yield an accurate poll if the bias is known and measurable, which is exactly what the next step exists to correct. What kills a poll is an invisible bias: a group that never appears in the sample at all, in proportions nobody can estimate. Good pollsters obsess over their sample frames; bad ones buy a cheap panel and hope, and the difference shows up on election night.

Weighting and the Likely-Voter Model: Where Polls Are Really Made

Weighting is the adjustment pass: if the sample contains too few voters without degrees, too many pensioners, or the wrong regional mix versus the known electorate, responses are mathematically rebalanced — some respondents’ answers count slightly more, others slightly less — until the sample matches official benchmarks for age, sex, region, past vote, and increasingly education. That last variable carries a famous lesson: in the 2016 US election, many state polls under-weighted voters without college degrees, a group that broke heavily one way; the samples literally contained too few of the voters who decided the outcome, and no amount of elegant modelling downstream could repair an upstream absence. Since then, education weighting has become near-universal. After weighting comes the poll’s most consequential judgement call: the likely-voter model. Elections are decided not by registered voters but by those who actually turn out, so pollsters screen and weight by intention (“certain to vote”), past voting history, and enthusiasm — and a model tuned for a high-turnout year will misread a low-turnout one, which is why two honest polls of the same race can differ by several points in the same week: they disagree about who the electorate is, a question only election day answers. When you read a poll, these are the lines of the methodology worth your thirty seconds: sample size, dates, population (adults, registered, or likely voters), and whether education weighting is applied. Everything else is presentation.

Reading a Poll in Ninety Seconds: A Citizen’s Checklist

Everything above compresses into a fast routine for the next poll you meet. Who conducted it, and who paid? A pollster with a public track record and a published method outranks a campaign’s internal numbers released selectively — campaigns publish polls the way anglers photograph fish, keeping only the flattering angles. When was it in the field? A poll conducted before a major debate, scandal, or announcement describes a world that no longer exists; check the dates against the news. Whom did it ask? “Likely voters” in the final fortnight is the gold standard; “all adults” in a midterm will flatter whichever side’s supporters vote more reliably. What exactly was asked? Question wording moves numbers by points — “support” versus “favour,” a candidate named with their party or without — and honest pollsters publish the full questionnaire; read it before the headline. How does it compare with the average? A single survey two points from every other poll is most likely noise, not news; averages of recent quality polls are the least wrong estimate available at any moment. Run those five checks and you will read election coverage with the calm of the professionals who make their living by it — who treat any individual poll the way sailors treat a single wave: real, measurable, and no guide at all to the state of the sea.

The Margin of Error — and the Real Reasons Polls Miss

The margin of error (typically ±3 points for a 1,000-person sample) is widely misunderstood in both directions. It means: if this exact poll were repeated many times, the result would fall within that band of the truth most of the time — for the share of each candidate, which means a two-point lead between candidates is really a gap that could plausibly be six points or minus two. It does not cover any mistake in sampling, weighting, or turnout modelling; it quantifies only the luck of the draw. Polls therefore miss for reasons the margin never promised to cover. Late swing: a poll is a snapshot of its fieldwork dates, and campaigns move in the final week — several famous “polling failures” were accurate polls of a race that then changed. Non-response bias: if one candidate’s supporters systematically refuse surveys at higher rates — through distrust of media or pollsters — the sample silently undercounts them in ways weighting cannot fix, because you cannot weight a group you cannot identify in the data. Herding: some pollsters, consciously or not, nudge methods toward the consensus, making the average look safer than any individual poll deserves. The mature way to consume polling: read averages rather than single surveys, treat leads inside the margin as ties, weight pollsters by their historical accuracy and transparency, and remember that a poll measures the race as it was — a forecast only in the hands of readers who insist on treating it as one.