Every election season produces the same ritual: two polls of the same contest appear on the same day, several points apart, and partisans declare one of them rigged. Usually neither is. Polling is measurement under uncertainty, and honest measurements of a divided electorate, taken with different methods, routinely differ by more than the headlines suggest. Understanding why is the difference between reading polls and being used by them.
Begin with the margin of error — and then go beyond it. The stated margin covers only sampling randomness for the full sample; it roughly doubles for subgroups, and it excludes every other error source: who refused to answer, how undecideds were allocated, whether the sample was weighted correctly. Analysts writing for realnewsanalysis.com stress that the margin of error is the floor of a poll’s uncertainty, not the ceiling.
Weighting and the likely-voter puzzle
Modern polling’s central difficulty is that the people who answer surveys differ from those who do not — by education, trust in institutions and political engagement. Pollsters correct with weighting, adjusting samples to match the expected electorate. Small weighting choices produce large result differences, and reputable firms disclose theirs. The Pew Research Center has documented how methods — live phone, online panels, text-to-web — now yield systematically different averages, which is why aggregates across many pollsters beat any single survey.
Layered on top is the likely-voter screen: the pollster’s model of who will actually vote. In low-turnout contests these models matter more than the questions themselves. Two firms can interview similar people and publish different numbers purely because one believes the electorate will be older than the other does.
How to consume polls like an analyst
Averages deserve one more caveat: they inherit shared errors. When every pollster misjudges turnout the same way — as has happened in several famous election misses — the average is precisely, confidently wrong together. The defence is diversity of method: an average that blends phone, online and text surveys, with different turnout models, is sturdier than a dozen polls cut from the same cloth.
And a final rule for the reader’s sanity: polls measure the race at fieldwork time, under the news of that week. Treating movement within the margin as momentum is the commentary industry’s favourite fiction. The honest sentence is usually the boring one — the race is close, the trend is slow, and the undecideds will decide it. Analysts who say that sentence tend to be right more often than those who narrate the noise.
The practices are simple to state: read averages, not single polls; check the sample size, dates and sponsor; distrust any result more than a few points from the consensus until a second firm finds it too; and remember that a poll is a snapshot with error bars, not a prophecy with a margin of embarrassment. The polls are rarely lying. They are answering slightly different questions, noisily — and the noise is the story.
