How polls work, when they have failed, what they do to campaigns, and whether they should be restricted. The spec pairs polls with media influence - and the strongest essays treat polling as both a barometer and a driver of politics. This walk-through covers the machinery, the record from 1970 to 2024, the campaign effects, and the regulation debate, with the exam application at the end. Three short quizzes break the tour up.
An opinion poll is a survey of voting intention or political attitudes across a representative sample, usually 1,000 to 2,000 respondents, aiming to estimate - within a margin of error - what the wider public would say if asked. In the UK polls are run by competing private companies (YouGov, Ipsos, Opinium, Savanta, Survation, More in Common and others), commissioned by newspapers, broadcasters, parties and academics, with the British Polling Council setting voluntary industry standards. The spec point is 4.2, the influence of the media - 'the importance and relevance of opinion polls, media bias and persuasion' - but polling evidence also serves democracy-and-participation and political-parties essays. One body of evidence, three essay homes.
Scroll - each part of the machinery lights beside you.
You do not need a statistics course - you need five working parts: sampling, weighting, margin of error, turnout adjustment, and MRP. Each one is also an exam point, because each one explains a famous failure or a famous success.
A poll takes a constructed sample, weights it to look like the population, discounts the people unlikely to vote, and publishes a share with a margin of error the headline usually drops. Every stage can go wrong - and each famous failure went wrong at a different stage.
Modern political polling uses quota sampling - the sample is built to match the population on age, gender, region and social grade - mostly through recruited online panels (YouGov, Opinium, Savanta). Phone polling is nearly extinct: expensive, slow, and suffering low response rates. The risk of online panels: under-representing offline, older and less politically engaged voters - exactly the bias the 2015 failure exposed.
Raw responses are weighted to match the population on demographics and on past vote - how people voted at the last election and in the EU referendum. Past-vote weighting was strengthened after 2015, when the Sturgis Inquiry found samples had over-represented engaged Labour voters. Weighting is the correction layer - and when the electorate does something genuinely new, the corrections built from the last election stop working.
A 1,000-person poll carries roughly a plus or minus 3% margin per party share, so a 1-point lead is not statistically a lead - a caveat press coverage routinely ignores. Add herding: pollsters' tendency to cluster their published numbers around a perceived consensus, which reduces the apparent spread between firms and also reduces the real information in any single poll.
Pollsters discount respondents unlikely to vote - and this is the source of most modern errors. Younger voters often say they will vote, then do not; sometimes (2017) they say they will and actually do. The pattern across the failures: the hard problem is not measuring what people say, it is predicting who turns out.
MRP (Multi-level Regression and Post-stratification) takes one large national sample, models the demographics of every constituency, and predicts each seat. YouGov pioneered it for UK politics in 2017 - and its MRP correctly predicted the hung parliament when conventional polls did not. It is now the standard final-week tool, alongside the joint BBC/ITV/Sky exit poll: 2019's predicted a Conservative majority of 86 (actual 80) and 2024's predicted a Labour majority of 170 (actual 174) - both highly accurate.
Scroll - each election lights with what the polls said and what happened.
The strategic point first, because it is the difference between a thin essay and a strong one: 'the polls were wrong in 1992 and 2015' is fine but weak on its own. The strong version pairs each failure with the methodological reform it produced - the Sturgis Inquiry after 2015, MRP from 2017 - to show that polling has improved and that the distortion now often sits in the media coverage of polls, not in the polls themselves.
1970, 1992 and 2015 are the failures; 2017 is the failure with one famous exception; 2019 and 2024 are the recoveries. The recurring cause: turnout - models built on the last electorate breaking when the next one behaves differently.
Polls showed Wilson winning comfortably; Heath won. The first major modern UK polling failure, and the opening of the long argument about reliability that has never really closed.
All major polls predicted a narrow Labour win or a hung parliament. The Conservatives won 41.9% to 34.4% - a majority of 21. The failure produced the 'shy Tory' thesis: Conservative voters under-reporting their intentions to interviewers. The classic case for any media-influence essay.
Polls showed a Conservative-Labour tie or a thin Labour lead; the Conservatives won an outright majority. The Sturgis Inquiry diagnosed sampling bias - the panels over-represented politically engaged Labour voters - plus weighting failures, and the industry overhauled its methodology. The failure-and-reform pairing starts here.
Most polls showed a comfortable Conservative majority throughout the campaign; the result was a hung parliament, with Labour gaining 30 seats on a late surge among younger voters that conventional turnout models missed. The exception: YouGov's MRP correctly predicted the hung parliament. 2017 is therefore evidence for both sides - polling can fail badly, and polling done with better method can succeed at the same election.
2019 polls were broadly accurate on the size of the Conservative lead - credibility partially restored. 2024 polls were accurate on the Labour lead, though several over-stated the majority and the Reform vote was harder to read than the Conservative collapse. The exit polls at both - majority of 86 predicted against 80 actual in 2019, 170 against 174 in 2024 - were highly accurate. The honest verdict for essays: improved, not infallible, and weakest when something genuinely new (Corbyn 2017, Reform 2024) breaks the turnout model.
Effects on campaigns, polls between elections, and whether late polls should be restricted.
Campaign effects. The bandwagon effect moves voters toward the perceived winner - some evidence in 1997, when months of landslide polling preceded a landslide result. The underdog effect moves them the other way - disputed, with some evidence that Labour's late 2017 surge was an anti-landslide rebound. Tactical voting is the clearest effect of all: the Lib Dem revival in the 2024 South-East depended heavily on tactical-voting polling and constituency-level MRP. Behind the scenes, parties use internal polling to decide which seats to defend and where to send their leaders, and the daily numbers drive the press narrative of who is winning.
Between elections. Most political polling lives between elections, not during them: voting-intention trackers, leader approval ratings, issue-salience tracking and by-election polling. Between-election polls work as a running check on government - and they drive the leadership-crisis cycle: when a leader's numbers drop sharply (Truss in September 2022, Corbyn from late 2018), rebellions and challenges become more likely. The polls do not cause the crisis, but they accelerate it - a strong democracy-and-participation point about polls as a constraint on prime ministerial power.
The ban debate. France, Italy and Singapore restrict published polls in the days before an election; the UK relies on British Polling Council self-regulation. The case for restriction: polls can mislead when wrong (1992, 2015), they can produce bandwagon or underdog distortions, and late polls drive last-minute coverage that may not reflect the actual mood. The case against: the free flow of information is itself a democratic value; the 2019 and 2024 record shows self-regulation broadly working; voters use polls to vote tactically, which is a legitimate strategy; and a ban only displaces polling into private hands - parties and well-connected journalists would still have the numbers, and ordinary voters would not.
Three essay homes, past questions, and the checklist for any polling reference.
Polling appears in Paper 1 Section A 30-mark questions (AO1/AO2/AO3 split 10/10/10) in three places. In Voting Behaviour and the Media, the failures (1992, 2015, 2017) evidence media-shaped narratives, and MRP 2017 shows polling done well. In Democracy and Participation, the regulation debate and polls-as-a-check-on-government are the angles. In Political Parties, between-election polling evidences party-system change and leadership pressure. The board has also set polls directly as a source question.
Past questions where polling evidence carries the answer.
Approach: The direct polls question - a source question, so pair the source's two views and analyse only its information, using your own knowledge to develop source points. Advantages: information for voters, tactical voting, a check on government, the accurate 2019 record. Disadvantages: 1992 and 2015 misleading voters, bandwagon and herding distortions, narrative-driven coverage. Land a clear line - the failure-plus-reform pairing supports 'more advantages, with reporting reform'.
Approach: Polls are your precision tool here. Poll-driven narratives (Corbyn's record-low ratings shaping 2019 coverage) show media power working through polling; the 1992 and 2015 failures show coverage amplifying wrong numbers. The counter: the 2024 Sun switch to Labour followed the polls rather than leading them - the press backs winners more than it makes them.
Approach: Use polls to test the claim. If media (including polling coverage) decided elections, the months of landslide polling in 1997 and 2024 would be cause rather than measurement - but the deeper drivers (the Winter of Discontent, Brexit, the mini-budget) sat outside media control. Judgement: the media amplifies and frames; the four-elections evidence says context and leadership decide.
Five rules for any polling reference in an essay. They turn a dropped statistic into evidence.
'YouGov's MRP' beats 'a poll'. Named sources score as precise knowledge; anonymous ones read as guesswork.
Polling is a time series - a number without a date is meaningless. 'In the final week of the 2017 campaign' does analytical work on its own.
'The exit poll predicted a majority of 170 against an actual 174' is evidence; 'the exit poll was about right' is an impression.
If your point depends on a small gap, say whether the gap exceeds the margin of error. If your point is about method, name it - MRP, online panel, past-vote weighting.
Never just drop the stat. Every figure should finish a sentence that starts with a claim - the figure is the proof, not the point.
Opinion poll. A survey of voting intention or political attitudes across a representative sample, usually 1,000-2,000 respondents, estimating public opinion within a margin of error.
British Polling Council (BPC). The industry's voluntary self-regulation body. The UK has no statutory regulation of polling.
Quota sampling. Building the sample to match population proportions on age, gender, region and social grade - the modern norm; true random sampling is rare.
Weighting. Adjusting raw responses to match the population on demographics and past vote. Past-vote weighting was strengthened after 2015.
Margin of error. Roughly plus or minus 3% per party share on a 1,000-person poll. A 1-point lead is not statistically a lead.
Herding. Pollsters clustering their published numbers around a perceived consensus - reducing both divergence and information value.
Turnout adjustment. Discounting respondents unlikely to vote - the hardest call and the source of most modern errors.
MRP. Multi-level Regression and Post-stratification: one large national sample modelled down to every constituency. Pioneered for UK politics by YouGov in 2017.
Exit poll. The joint BBC/ITV/Sky election-day poll. 2019: predicted majority 86, actual 80. 2024: predicted 170, actual 174.
1970 failure. Polls said Wilson; Heath won - the first major modern UK polling failure.
1992 failure and the shy Tories. Narrow Labour win predicted; Conservative majority of 21 (41.9% to 34.4%). Conservative voters under-reported their intentions.
2015 failure and the Sturgis Inquiry. Tie predicted; Conservative majority delivered. Diagnosis: sampling bias over-representing engaged Labour voters, plus weighting failures. The industry reformed in response.
2017 - failure with an exception. Conventional polls missed the hung parliament; YouGov's MRP called it.
Bandwagon effect. Voters moving toward the perceived winner - some evidence in 1997.
Underdog effect. Voters moving toward the perceived loser - disputed; some evidence in the 2017 late surge.
Tactical voting. Using polls to back the best-placed challenger - central to the Lib Dems' 2024 South-East gains.
The ban debate. France, Italy and Singapore restrict late published polls; the UK relies on BPC self-regulation. The strong landing point: reform reporting standards rather than ban publication.