Explainer

Opinion Polls Explained: Sampling, Margin of Error & Why Polls Can Be Wrong

Opinion polls are the core data product of political science — but they are widely misunderstood. A single poll is a snapshot, not a prediction. Understanding how polls are made, what they can reliably tell us, and where they fail is essential to interpreting them intelligently.

What is an opinion poll?

An opinion poll estimates what a population thinks by measuring the views of a sample. The key insight from statistics is that a well-drawn sample of around 1,000 people can give a reliable estimate of what 47 million UK adults think — but only if the sample is genuinely representative.

Getting a representative sample is the hard part. The challenge is not just sample size; it is that people who agree to be polled are systematically different from people who refuse. Older people, those with stronger political views, and those with higher education are all more likely to respond. Left unaddressed, these biases distort results. This is why pollsters apply weighting to their raw data.

Random probability sampling

The gold standard of survey methodology is random probability sampling: every member of the population has a known, non-zero probability of being selected. In a true random probability poll, interviewers cannot substitute an easy respondent for a hard-to-reach one. The British Election Study uses probability sampling, but most commercial political polling uses quota methods instead because of the cost and time involved.

Quota sampling and online panels

Most UK political polls use quota sampling via online panels. Interviewers — or an online platform — are given targets to fill: X% women, Y% aged 18–34, Z% in each region. As long as quotas are met, it does not matter which specific individuals respond.

YouGov, Ipsos, and others recruit respondents from pre-existing online panels of hundreds of thousands to millions of members who take surveys for small rewards. Online panels have known limitations: panel conditioning (frequent survey-takers become more politically engaged than average), self-selection bias, and education over-representation. These issues are manageable with weighting but never fully eliminated.

Weighting

After collecting raw data, pollsters apply weights to make the sample match the target population. Common weighting targets in UK polling include age, gender, region, education, and — crucially — past vote recall. Past vote recall is how the respondent says they voted in the last election. Pollsters weight their samples so that the proportion saying they voted Conservative, Labour, etc. in 2024 matches the actual 2024 result.

This creates its own problem: people misremember how they voted (“false recall”), and the mix of who can be recruited shifts over time. There is no perfect weighting solution, which is why different pollsters applying different weighting schemes get different results — the so-called house effects.

House effects

House effects are the systematic tendencies of individual pollsters to produce results that consistently differ from the consensus in one direction. One firm might show Labour 3 points higher than average; another might show Reform 4 points higher. House effects arise from weighting choices, sampling frames, likely voter models, question wording, and how undecided voters are handled. Our house effects explainer documents the current bias of each major pollster.

Margin of error

Every poll has a margin of error — the range within which the true figure lies with a given level of confidence. For a simple random sample of 1,000 people, the margin of error is approximately ±3 percentage points at 95% confidence. A poll showing Labour at 18% and Reform at 28% should be read with that uncertainty in mind, though the 10-point gap is well outside any plausible margin of error.

Important caveat: the formal margin of error applies to random probability samples. For quota samples, the real uncertainty is somewhat larger because of additional sources of error beyond pure sampling variation.

Herding

Herding occurs when pollsters adjust their results to be closer to the consensus, consciously or unconsciously. If a firm gets a result showing the Conservatives at 35% when everyone else shows 28%, they may scrutinise their methodology more carefully and make adjustments. The problem is that if all the consensus polls are wrong in the same direction, herding makes the collective error invisible. Herding was a major factor in the UK polling industry’s 2015 failure.

Why polls can be wrong: a brief history

• 1992: All polls predicted a hung parliament or slight Labour lead. Conservatives won by 7.6%. “Shy Tory” effect blamed — Tory voters reluctant to admit their preference
• 2015: Polls consistently showed a near-tie. Conservatives won by 7 points. A BPC/MRS inquiry found herding and non-representative samples as the main culprits
• 2017: Polls understated a late Labour surge. Conservatives expected to win easily; they lost their majority. See 2017 results
• 2024: Polls correctly predicted a large Labour win; the scale was broadly right. See 2024 accuracy table for how every pollster performed

How to read a poll intelligently

• Never read a single poll in isolation. Look at the polling average
• Check the fieldwork dates: a poll conducted over five days last week is less current than one completed yesterday
• Note the sample size: smaller samples (<500) have wider margins of error
• Know your pollster: different houses have different house effects. Our pollster profiles document these systematically
• Do not over-interpret small movements: a 2-point change in a single poll is likely noise, not signal
• Remember that VI polls measure intentions, not predictions: FPTP converts votes into seats in highly non-linear ways

The British Polling Council sets standards for how member pollsters conduct and report their work, including requirements to publish full methodology and data tables.