Explainer

Polling Methodology Explained: How UK Polls Are Conducted, Weighted and Tested

Understanding how UK opinion polls are made is essential for interpreting them correctly. A poll is not a prediction — it is a snapshot of a sample. The reliability of that snapshot depends on how the sample was drawn, how it was weighted, and whether the methodology has been rigorously tested against real election results.

Online, telephone and MRP: the three main methods

UK political polling uses three main data-collection methods, each with distinct trade-offs between cost, speed, and accuracy.

Online panel surveys are by far the most common. Polling firms maintain large panels of adults who have agreed to take surveys, typically for small incentives. For each poll, the firm samples from its panel, recruits respondents to fill demographic quotas, and collects responses in hours rather than days. Online panels are fast and cheap — a 1,000-person survey can be completed overnight — but suffer from systematic biases: panel members tend to be more politically interested, more educated, and more digitally active than the population at large.

Telephone surveys once dominated political polling but have declined sharply. The shift from landlines to mobile phones destroyed the sampling frames that made telephone probability sampling viable. Random digit dialling on mobile numbers is expensive and produces very low response rates. A small number of firms still conduct telephone polls, and some evidence suggests they produce slightly different results to online methods — often showing smaller leads for populist parties whose supporters are harder to recruit online.

MRP (Multilevel Regression and Post-stratification) is a fundamentally different approach. Rather than a single national survey, MRP combines a large-scale poll of 10,000 to 50,000 respondents with detailed demographic data for each of the 650 Westminster constituencies. A statistical model estimates vote intention at constituency level, then post-stratifies to match the known demographic profile of each seat. MRP is expensive but significantly more accurate for seat projections. YouGov’s MRP correctly predicted a hung parliament in 2017 when other polls were pointing to a large Conservative majority.

Sample sizes and what they mean

A standard UK political poll uses between 1,000 and 2,000 adults. The formal margin of error for a random sample of 1,000 is approximately ±3 percentage points at 95% confidence. This means if a poll shows Reform at 29%, the true figure could plausibly be anywhere from 26% to 32%.

Three important caveats apply to this figure. First, the margin of error assumes a pure random probability sample — most UK polls use quota sampling, which has additional sources of error. Second, the margin of error applies to each individual party estimate independently, not to the gap between parties. Third, small subgroup estimates — such as voting intention among 18–24 year olds — have much larger uncertainty because the subgroup is only a fraction of the full sample.

MRP models use 10,000 to 50,000 respondents nationally, but individual constituency estimates are still based on small numbers of actual respondents from that seat. The accuracy of constituency-level MRP projections comes from the statistical model, not from large constituency samples directly.

Weighting and its limits

After data collection, pollsters apply weights to correct for sample imbalances. A respondent who fills an underrepresented demographic quota cell (e.g. a young working-class man without a degree) receives a higher weight; an overrepresented respondent receives a lower weight. Common weighting targets in UK polls include age, gender, region, social grade, and educational attainment.

The most controversial weighting variable is past vote recall: respondents are asked how they voted at the previous general election, and the sample is weighted so that the recalled 2024 vote matches the actual 2024 result. This corrects for the well-documented tendency of winners’ voters to be more easily recruited. The problem is that past vote recall is unreliable — people misremember, misrepresent, or simply did not vote — and the pool of respondents who can be recruited at all shifts over time, making it impossible to perfectly reconstruct the electorate from 2024 in a 2026 survey.

Different weighting choices by different firms produce what are called house effects: systematic tendencies to show one party higher or lower than the cross-pollster average. A polling average that combines multiple firms cancels out some of these house effects.

Herding: the hidden danger

Herding occurs when pollsters adjust their results toward the industry consensus, consciously or unconsciously. If a firm produces a poll showing the Conservatives at 25% when every other firm shows 19%, it will subject its methodology to extra scrutiny. Sometimes that scrutiny finds a genuine error; but sometimes the result was correct and the consensus was wrong. Either way, the incentive is to converge.

Herding makes polling errors systematic rather than random. If all firms converge on the same wrong answer, the error is invisible until election night. The 2015 UK polling miss was the clearest example: every firm showed a near-tie; the Conservatives won by seven points. The British Polling Council and Market Research Society inquiry found clear evidence of herding alongside unrepresentative samples.

Detecting herding is straightforward statistically: if the variance between polls from different firms is smaller than would be expected by chance given their sample sizes, herding is likely occurring. This test is routinely applied to UK polling data and shows clear clustering in some election cycles.

Historical accuracy: 2015 to 2024

UK polling has a mixed accuracy record over the last three general elections:

• 2015: major failure. All polls showed a near-tie between Conservative and Labour. The Conservatives won by 6.5 points. The BPC/MRS inquiry identified unrepresentative online samples and herding as the primary causes. Pollsters were producing samples skewed toward Labour supporters, then converging on each other’s wrong estimates.
• 2017: partial miss. Most polls correctly predicted a Conservative lead, but underestimated a late Labour surge. The Conservatives were forecast to win a landslide majority; they lost their majority entirely. YouGov’s MRP correctly predicted a hung parliament; standard polls did not capture the late swing to Labour, possibly because fieldwork was completed before the final days of the campaign.
• 2019: broadly correct. Polls correctly pointed to a significant Conservative lead, and most firms gave the Conservatives a double-digit lead. The actual result was a 12-point Conservative margin. Individual firms varied considerably, with some overestimating and some underestimating the scale of the win, but the directional call was right across the industry.
• 2024: close but not perfect. Polls correctly predicted a large Labour victory. The actual Labour vote share (33.7%) was close to polling averages, though Reform UK’s 14.3% final vote share was somewhat below late polling averages of 17–18%, suggesting some late switching or differential turnout. Seat projections varied widely because translating vote shares to seats under FPTP involves substantial model uncertainty.

What good polling literacy looks like

A single poll tells you relatively little. The margin of error alone allows for a range of outcomes that spans most plausible political scenarios. The value of polling comes from aggregation: comparing results across firms, looking at trends over time, and weighting by methodological quality.

Key habits for reading polls intelligently:

• Always check the sample size and fieldwork dates, not just the headline figures
• Compare against the cross-pollster average, not against the previous poll from the same firm
• Be sceptical of large single-poll swings; look for confirmation across multiple firms before treating a move as real
• Distinguish between voting intention questions (hypothetical ballot tomorrow) and actual election forecasts
• Know which firms tend to show higher or lower figures for which parties — house effects are real and persistent

Frequently Asked Questions