United Kingdom General electionby 15 Aug 2029in 999 days Majority for any party 34% Labour largest party 93% Calendar Track record Research
Research · United Kingdom · Foundations

From answers to headlines: the stage model of a published British poll

Abstract. A published voting-intention figure is the end of a chain of steps (weighting, a turnout filter or weight, the handling of don't knows), and each pollster runs the chain differently. We store every poll's tables at each stage, derive England from the GB banner by one of four stated rules, check the arithmetic automatically, and find that at 2015–2024 final polls the answers before those steps missed in a steadier direction than the headlines did.

Polling Forecast · updated 9 November 2026 · 21 min read · poll methodology, turnout, don't knows, weighting, England, data tables, raw sample

What has changed since this paper was written

Updated 9 November 2026. The rest of the paper is as first published, except where a sentence below says it was corrected.

  • Polls without tables (26 October). GB polls taken from Wikipedia (grade C) count in the GB average only. They have no tables, so they have no raw stages and play no part in the stage model until the pollster's own tables replace them.
  • Names on the site (3 and 8 October). "Our reading" is shown as our adjusted average, and the headline figure as our projection.

A headline voting-intention figure looks like a direct measurement: 30% of people say they will vote Labour. It is not. Between the answers people give and the figure that is published, every pollster applies several steps. The sample is weighted to resemble the population, usually including how people say they voted last time. Some respondents are set aside or down-weighted because they seem unlikely to vote. People who say they don't know are asked again, modelled, or dropped. Each step is defensible, each is a judgement, and each pollster makes it differently.

This paper describes how we record those steps. We call it the stage model: every poll's published tables are stored as printed, and each table is assigned to a stage in the chain from answers to headline. The model has three uses. It makes the pollsters' adjustments visible, poll by poll, on every poll page. It lets us derive a consistent England figure from tables designed for Great Britain. And it gives us a second measurement of polling error, at the stage before the pollsters' own adjustments, which turns out to behave differently from the headline error measured in Measuring British polling error in England, 2015–2024. That second measurement is the basis of our reading of the polls, set out in Correcting the raw answers: the method behind our UK projection and Why we start from British pollsters' raw answers rather than the published headline.

The stages

Each table read from a pollster's file is given one of six stage codes. Four of them form a sequence.

S0, unweighted
The answers as collected, before any weighting, including don't know, would not vote and refused. Only Ipsos prints this among the pollsters we read (its "UnW Tot." column).
S1, weighted, everyone asked
The answers after the pollster's demographic and political weighting, for everyone asked the question (or the widest base the pollster prints), with don't knows and non-voters still included. No turnout step has been applied.
S2, after the turnout step
The same answers after the pollster's turnout filter or turnout weight, still including don't knows. Some pollsters print this table; others go straight from S1 to the headline.
S3, headline
The published voting intention: parties only, after turnout and the treatment of don't knows.

Two further codes hold tables that feed the chain but are not points along it.

SQUEEZE
A follow-up question put only to those who said they didn't know, typically "if you had to choose, which party would you vote for?"
LTV
The likelihood-to-vote question, on whatever scale the pollster uses (0–10, 1–10, or a four-point verbal scale).

Every table also carries a base scope, which records who it covers. The scope matters because two S1 tables can describe different people.

Base scope Who is in the table Example
AllAdults Everyone asked Find Out Now "Voting"; Ipsos "Combined VI – all"
ExcludesLtv0 Everyone except those certain not to vote Survation Table V2_1 (excludes 0 out of 10); Opinium V003 "who might vote"
Ltv5Plus Those at 5 or more out of 10 Defined for completeness; no current reader uses it
LikelyVoters Those the pollster classes as likely to vote Ipsos S2 (9–10 out of 10); Opinium V003 in 2025 files worded "likely to vote"
DontKnowsOnly Those who said don't know at the first question Every SQUEEZE table
HeadlineBase Whoever the pollster's headline counts Every S3 table
Other A comparison table outside the main chain YouGov's standard question (see below)

Table: Base scopes used in the stage model.

Each stored table keeps its full banner, not only the Total column: age, sex, region, nation, 2024 vote and whatever else the pollster prints. Each column carries its printed weighted and unweighted base. Every table records the source file's name and SHA-256 hash, the sheet or page, the reader and its version, and the results of the arithmetic checks described below. Nothing is updated in place. A table extracted again, for example by a corrected reader, is inserted as a new row and the old row is marked as superseded, so every figure on the site can be traced to the exact file it came from. The original files are also kept in the database.

For comparison across stages we use a parties-only form. If \(\pi_a\) is the share of answer \(a\) in a table (parties, don't know, would not vote, refused), and \(P\) is the set of parties, the parties-only share is

\[ s_a = \frac{\pi_a}{\sum_{b \in P} \pi_b} \times 100, \qquad a \in P. \tag{1} \]

Equation (1) applied to an S1 table gives what we call the raw answers: weighted, everyone asked, don't knows set aside. The Labour lead at any stage is \(s_{\text{LAB}} - s_{\text{CON}}\) on this basis.

How six pollsters get from S1 to S3

We hold full tables for six pollsters: Find Out Now, More in Common, Survation, Ipsos, YouGov and Opinium. Each has its own reader, a C# class written for that pollster's file format. The readers find tables by their labels (sheet names, table codes, question text, the word "Total" in a header row), never by fixed positions, so that small layout changes are caught rather than misread. A reader that meets a row label it does not recognise stores it as unknown and reports it, rather than guessing.

Each reader also records the pollster's method as stated in its tables or methodology notes: mode, weighting targets, whether it weights to past vote, how it handles turnout and don't knows, whether its question names the respondent's constituency, and whether its headline is the output of a statistical model. The table below summarises what is stored.

Pollster Tables S1 (raw) Turnout step Don't knows Weighted to past vote Question
Find Out Now Excel, weekly Everyone, weighted; don't know included Filter: "definitely" or "very likely" to vote (2024 non-voters only if "definitely") Squeeze; those still undecided excluded Yes, 2024 vote (with gender, age, region, ethnicity) Names the constituency in files since spring 2026; standard wording before
More in Common Excel, weekly Everyone, weighted; would not vote and don't know included Filter: 8 or more out of 10 (9 or more for 2024 non-voters) Squeeze, then a statistical model that is not tabulated Yes, 2024 vote (with age and sex, region, ethnicity, education) Standard
Survation Excel Everyone except 0 out of 10 Weight: answers multiplied by likelihood to vote ÷ 10 (S2 printed) Undecided and refused removed Yes, 2024 vote and 2016 EU referendum vote Names the constituency
Ipsos PDF All adults, weighted; S0 also printed Filter: 9–10 out of 10 and always, usually or "it depends" vote (S2 printed) Squeeze; those still undecided excluded No: 2024 vote used to stratify the sample, not to weight it Standard
YouGov PDF, weekly Everyone, weighted, constituency question Inside an MRP model; not visible Modelled inside the MRP Yes, 2024 vote by region (with political attention and others) Names the constituency
Opinium Excel, fortnightly Those who "might vote" (base filter) None beyond the base filter Don't know and would not vote removed Tables say "politically representative"; recorded as past-vote weighting Standard

Table: How each pollster's published figure is produced, as recorded by our readers. Sources: each pollster's cover sheets, table bases and methodology notes.

A few features of the table need more explanation.

Find Out Now prints four sheets: "Voting" (S1), "Squeeze", "Turnout" (a four-point likelihood scale, stored as LTV) and "Headline VI" (S3). It does not print the stage between the turnout filter and the squeeze, so the chain for this pollster is S1 to S3 in one step. Its S1 question has no "would not vote" option, which probably explains why its don't-know share is much higher than other pollsters'.

More in Common likewise goes from its "votingintention (raw)" sheet (S1) to its "votingintention (headline)" sheet (S3), with a 0–10 likelihood table alongside. The final assignment of remaining don't knows uses a model that More in Common describes on its methodology page but does not tabulate, so that part of its step cannot be reproduced from the files. Its recent files print only unweighted bases, which affects how England is derived (next section).

Survation is the most transparent about turnout. Table V2_1 is the full question for everyone except those who say they are certain not to vote. Table V2_2 is the same answers weighted by stated likelihood to vote divided by ten, so someone at 10 out of 10 counts fully and someone at 5 counts half. Table V2 removes the undecided and refused and is the headline. Survation's totals are UK-wide, with Northern Ireland as a banner column.

Ipsos is read from PDF tables. "Combined voting intention – all" gives S1, and its unweighted total column gives S0, the only unweighted stage any of these pollsters publishes. "Combined voting intention – likely to vote" gives S2, and the party rows of that same table, divided by its "Total voting" sub-base, give the headline. The reader stores the weighted counts rather than the rounded percentages, and checks every printed percentage against them. Ipsos is also the only one of the six that does not weight to past vote. Its sample comes from a probability-based online panel (KnowledgePanel).

YouGov has published its headline through a multilevel regression and post-stratification (MRP) model since 2025, so its S3 is flagged as a model output. Turnout and don't knows are handled inside that model and cannot be seen as separate steps. Its S1 is the question that names the parties standing in the respondent's own constituency, because the model is built from that question. YouGov also prints the standard question ("if there were a general election held tomorrow…"); we store it as a comparison table with base scope Other, and wherever the site needs a raw stage it prefers the non-Other table. YouGov prints whole-number percentages throughout.

Opinium is the simplest case. Its full question (V003) is asked of people who might vote, and its headline is that table with don't know and would not vote removed. There is no turnout weight between them, so its raw answers and its headline agree on the parties. Its turnout step, in effect, is the base filter. Two early files lack V003 and only the headline is stored; two February 2026 files print a headline without the SNP and Plaid Cymru rows, which does not add up to 100, so that headline is left out and V003 kept.

The differences are not cosmetic. Three pollsters filter on stated likelihood, one weights by it, one models it, and one does neither beyond its base. Two assign remaining don't knows in ways that cannot be reproduced from the tables; the others drop them. Three ask about the respondent's constituency and three do not. Averaging headlines from these six treats them as measurements of the same thing. The stage model at least makes clear where they differ.

What each pollster publishes is scored separately, as part of the pollster ratings: whether it prints a raw stage, a turnout stage, a past-vote table, a likelihood-to-vote table and an England figure (see Rating the British pollsters: accuracy, the Labour–Conservative lead and openness).

Deriving England from a GB banner

Almost every pollster samples Great Britain, or the United Kingdom, and publishes England only as a column or a set of regions within a GB table. The England figure must be derived from that banner. The rules are applied in a fixed order, and the method used is recorded with every figure.

Let \(x_{a,c}\) be the percentage for answer \(a\) in banner column \(c\), and \(w_c\) that column's printed weighted base.

Column. If the pollster prints an England column, it is used as printed:

\[ \pi^{E}_{a} = x_{a,\text{ENG}}. \tag{2} \]

RegionSum. Otherwise, if weighted bases are printed and every English region is present, England is the base-weighted average of the regions:

\[ \pi^{E}_{a} = \frac{\sum_{r \in R} w_r \, x_{a,r}}{\sum_{r \in R} w_r}, \tag{3} \]

where \(R\) is either the nine ONS regions (North East, North West, Yorkshire and the Humber, East Midlands, West Midlands, East of England, London, South East, South West) or the four coarse regions some pollsters use (North, Midlands, London, South). This is exact, in the sense that it reproduces what an England column would show, provided the regions partition England and the bases are the weighted bases of the same table.

TotalMinusNI. This rule gives GB rather than England. When a pollster's totals cover the UK, Northern Ireland is removed using weighted counts:

\[ \pi^{GB}_{a} = \frac{n_{a,T} - n_{a,\text{NI}}}{w_T - w_{\text{NI}}} \times 100, \tag{4} \]

where \(n_{a,c}\) is the weighted count of answer \(a\) in column \(c\) (printed where the pollster gives it, otherwise \(x_{a,c} w_c / 100\)), and \(T\) is the Total column. Survation's and Opinium's GB figures are derived this way; Opinium's Northern Ireland column is empty, so the subtraction changes nothing but the rule is the same.

PopulationShares. When no weighted bases are printed, as in More in Common's recent files, the regions cannot be combined by equation (3). England is then taken as GB minus Scotland and Wales at their shares of GB adults:

\[ \pi^{E}_{a} = \max\!\left(0,\; \frac{\pi^{GB}_{a} - \alpha_S \, x_{a,S} - \alpha_W \, x_{a,W}}{1 - \alpha_S - \alpha_W}\right), \qquad \alpha_S = 0.086,\; \alpha_W = 0.049. \tag{5} \]

The constants are the shares of GB adults aged 18 and over living in Scotland (8.6%) and Wales (4.9%), from the ONS mid-2023 population estimates, so England's share is 86.5%. Pollsters weight their regions to population shares of this kind, which is why the rule works. It is an approximation, for two reasons. The headline base is a filtered subset, and the share of it in Scotland and Wales need not equal their share of adults. And percentages printed to finite precision lose a little information. The floor at zero stops small negative values for parties that do not stand in England. England figures derived this way are flagged, and they get a wider tolerance in the checks below.

The rule was validated before it was adopted. For polls where the regional weights could be bounded by linear programming over every set of weights that reproduces the GB headline, the population-share figure agreed with the bounded figure to within 0.2 points on 15 polls.

As an illustration, More in Common's poll of 18–21 September 2026 prints a GB headline of Labour 26.32%, with Scotland at 18.50% and Wales at 32.78%. Equation (5) gives England Labour as \((26.32 - 0.086 \times 18.50 - 0.049 \times 32.78)/0.865 = 26.74\%\). The same calculation for the Conservatives (GB 23.00%, Scotland 19.34%, Wales 20.34%) gives 23.52%.

Automatic checks

Every table is checked when it is read. The checks are arithmetic: they test whether the numbers are internally consistent and whether the reader has understood the table, not whether the poll is right. A file whose checks fail is not stored. When a file arrives through the daily automatic check, it is stored only if every check passes and the poll has no tables from a different file already; anything else waits on the admin page for a person.

Check What it tests Tolerance
Total sums to 100 The answers in the Total column add up to 100% ±1.5 points; ±1.0 for Ipsos headline parties; ±3.0 for YouGov's whole numbers
Counts match shares Each printed weighted count divided by its column base matches the printed percentage Per cell, \((50 + 0.5\,x)/w + 0.15\) points, allowing half a respondent in the count and rounding in the base; Ipsos's whole-number percentages get a further 0.5
England matches regions Where an England column and English regions are both printed, equation (3) reproduces the column 1.0 point on the worst answer; 2.0 for YouGov
Base is the sample The weighted base of the all-respondents table equals the sample size stated on the cover ±1 respondent (Find Out Now, Ipsos, YouGov)
Headline matches database The England headline rebuilt from the S3 table matches the England figures already stored for the poll 0.25 points; 0.6 for PopulationShares; 0.75 for whole-number tables

Table: Checks run on every table, with their tolerances. \(x\) is the printed percentage and \(w\) the column's weighted base.

The counts check has one variant. Some Survation releases print proportions to only two decimal places, in which case the weighted counts are more precise than the percentages. The reader then uses the counts as the values and relaxes the counts check by half a point.

The last check links the stage tables to the rest of the database. Every poll in the England average has England figures in the main poll results table, entered and checked when the poll was first recorded. The headline rebuilt from the stored tables must agree with them. When the stage tables were first loaded, all 65 polls from Find Out Now, Survation and More in Common matched: Find Out Now and Survation to within 0.2 points, and More in Common's population-share England figures to within 0.2 points. Nine of Ipsos's ten releases in this parliament were read, all within 0.06 points; the tenth is a working draft with notes inside the cells and is refused with a message. All 90 YouGov files matched, the worst by 0.47 points because of whole-number rounding. All 54 Opinium files matched to within 0.001 points. A poll new to the database has no stored figures to compare against, so for it this check does not run; instead the poll is created from its own headline table, and only readers whose tables reproduce the headline exactly are allowed to do that.

A worked example: Survation, 9–10 September 2026

Survation's poll with fieldwork on 9–10 September 2026 (sample 2,043 UK adults) shows all three stages, which makes it a good example. The file is stored with 13 checks, all passed. The England column is printed, so equation (2) applies throughout.

At S1 (Table V2_1, everyone except those certain not to vote), the England column has a weighted base of 1,577 and shows:

England Con Lab LD Reform Green Restore Other Don't know
S1: printed share (%) 20.29 28.28 10.00 22.06 8.30 1.55 0.64 8.88
S1: parties only, eq. (1) 22.27 31.04 10.97 24.21 9.11 1.70 0.70
S2: printed share (%) 21.07 28.10 10.46 22.66 8.11 1.63 0.58 7.39
S2: parties only, eq. (1) 22.75 30.34 11.29 24.47 8.76 1.76 0.63
S3: headline (%) 22.75 30.34 11.30 24.47 8.76 1.76 0.63

Table: Survation, fieldwork 9–10 September 2026, England column at each stage. Printed values are proportions to four decimals, shown here as percentages.

Checking S1. The printed weighted count for Labour in England is 446. Dividing by the base, \(446/1577 = 28.28\%\), which matches the printed proportion of 0.2828 to within 0.002 points; the allowance for this cell is \((50 + 0.5 \times 28.28)/1577 + 0.15 = 0.19\) points. The printed answers sum to 100.0%.

S1 to raw answers. The parties together account for \(100 - 8.88 = 91.12\%\) of the S1 England base. By equation (1), Labour's raw share is \(28.28/91.12 \times 100 = 31.04\%\) and the Conservatives' is \(20.29/91.12 \times 100 = 22.27\%\). The raw Labour lead is \(+8.77\) points.

S1 to S2. Survation's turnout step weights each respondent by stated likelihood to vote divided by ten. In England this raises the Conservatives from 20.29% to 21.07% of the base and Reform from 22.06% to 22.66%, lowers Labour slightly and cuts the don't-know share from 8.88% to 7.39%: people who say they don't know are, on average, less sure they will vote. On a parties-only basis the Conservatives move from 22.27 to 22.75 and Labour from 31.04 to 30.34. The Labour lead falls from \(+8.77\) to \(+7.59\).

S2 to S3. Survation's headline removes the undecided and refused. The parties together make up \(100 - 7.39 = 92.61\%\) of S2, and dividing each party's S2 share by 0.9261 reproduces the printed headline (Labour \(28.10/0.9261 = 30.34\)). So for this pollster the don't-know step is exactly equation (1) applied to S2, and changes nothing on a parties-only basis. The whole of the movement between raw answers and headline, 1.2 points off the Labour lead, comes from the turnout weight.

The headline check. The England headline rebuilt from Table V2 agrees with the England figures stored for the poll (Labour 30.34, Conservative 22.75) to within 0.01 points, well inside the 0.25-point tolerance.

GB. Survation's totals include Northern Ireland, whose column in the headline table is entirely "other" (the Northern Ireland parties): 57 of the 1,729 weighted respondents. By equation (4), GB Labour is \(484/(1729 - 57) \times 100 = 28.95\%\), against 27.98% in the UK-wide Total column.1

Finally, the poll page shows a fourth row, our reading: the same raw answers with half the average raw-answer miss taken off each party, and rescaled. That step is described in Correcting the raw answers: the method behind our UK projection.

The historical stages dataset

The stage model would be of limited use if it covered only current polls, because there would be nothing to compare the raw answers with. So the final polls from 2015 to 2024 were loaded into the same framework (script 09 in the database build). The historical dataset stores England figures at two stages only, S1 and S3, since that is what can be recovered consistently from old tables: 329 figures for 19 final polls. Each figure has a file, page or sheet, and table reference behind it. One poll (Lord Ashcroft's 2015 poll) has only its S1 stage, because its England figures are not on the headline basis, so the comparison below uses the 18 polls that have both stages, across 2015, 2017, 2019 and 2024.

This set overlaps with the primary series of Measuring British polling error in England, 2015–2024 but is not identical. It includes Number Cruncher's 2019 and 2024 polls, whose England reconstructions are on an estimated basis and are therefore outside the primary series; and it does not include YouGov's 2024 poll, whose raw stage is not stored. The headline errors in brackets below therefore differ slightly from the primary table.

For each election \(t\) the raw error of party \(p\) is the average of the polls' parties-only S1 shares minus the England result, and the headline error is the same for S3, over the same polls:

\[ e^{\text{raw}}_{p,t} = \frac{1}{n_t}\sum_{i} s^{S1}_{i,p,t} - v_{p,t}, \qquad e^{\text{head}}_{p,t} = \frac{1}{n_t}\sum_{i} s^{S3}_{i,p,t} - v_{p,t}. \tag{6} \]
ElectionPolls with tablesConLabLDReformGreen
20152−6.6 (−5.7)+1.6 (+1.5)+2.0 (+1.8)+0.9 (+0.8)+1.1 (+0.9)
20174−2.1 (+0.6)−1.5 (−3.9)+0.7 (+0.7)+2.1 (+2.1)+1.0 (+0.6)
20196−3.3 (−1.6)+1.8 (+1.0)−0.5 (−0.5)+1.1 (+0.8)+0.4 (0.0)
20246−4.0 (−3.0)+6.1 (+4.8)−1.7 (−1.3)+1.0 (+1.3)−0.4 (−0.6)
Mean−4.0 (−2.4)+2.0 (+0.9)+0.1 (+0.2)+1.3 (+1.3)+0.5 (+0.2)
Live Raw answers (weighted, everyone asked, don't knows removed) minus the result, England, points; in brackets the published headline of the same polls. Computed from the database on 19 November 2026.

The comparison is the most important empirical result in this paper, and it is worth reading carefully. Averaged across the elections, the raw answers' Conservative error was −4.0 points and their Labour error +2.0. The published headlines of the same polls, in brackets, were closer on both.2 So the pollsters' adjustments, taken together, moved the figures in the right direction on average: they lowered Labour and raised the Conservatives.

But the direction of the raw error is steadier than the direction of the headline error. In the table as it stands, the raw answers understated the Conservatives at every election and overstated the Reform family at every election, while the headline's Conservative error changed sign in 2017. The pollsters' own steps moved Labour down at every election, which was the right direction in three and the wrong direction in 2017, when Labour did better than the raw answers suggested. An error with a consistent sign can be partly corrected from its history; an error whose sign flips cannot. That is the argument developed in Why we start from British pollsters' raw answers rather than the published headline, and it is why our reading starts from S1 rather than S3.

A single 2017 poll shows the mechanism. ComRes's final poll (the firm is now Savanta) had, in its widest printed table (weighted, likely voters), Conservatives 37.61% and Labour 37.30% of all answers, with 7.5% don't know and 1.9% refused. On a parties-only basis that is Conservatives 41.55%, Labour 41.21%, a Conservative lead of 0.3 points. ComRes's published headline had the Conservatives on 46.23% and Labour on 35.94%, a lead of 10.3. The England result was Conservatives 45.50%, Labour 41.98%, a lead of 3.5. The raw answers were much closer on Labour; the headline was closer on the Conservatives; and the lead was badly wrong in both, in opposite directions. Whatever ComRes's steps from likely voters to headline did, in that election they moved the lead the wrong way, and by a great deal.

The 2024 row shows the method's limit. Labour's England error was +6.1 points in the raw answers, and the headlines of the same polls, in brackets, were only somewhat better. The pollsters' adjustments helped, but not nearly enough, and nothing in the raw-answer record from 2015 to 2019 anticipated a Labour overstatement of that size.

When the stage tables for current polls were first loaded, the three pollsters then read moved the England Labour lead towards the Conservatives by 0.6 (Find Out Now), 1.3 (Survation) and 1.4 (More in Common) points on average between raw answers and headline. That is small next to the Labour lean of the raw answers at past final polls, which ranged from under one point to over ten across pollsters and elections. The pollsters are adjusting in the historically right direction, but by less than history suggests they needed to.

What "raw" means, and what it does not

The word "raw" is convenient and misleading, so it needs a careful definition. In this project it means S1 on a parties-only basis: weighted, everyone asked (or the widest base printed), don't knows set aside, before any turnout step. It does not mean unweighted.

Raw answers are weighted. Every pollster we read except Ipsos weights its S1 table to recalled 2024 vote as well as to demographics. Weighting to past vote is itself a strong adjustment. It corrects a sample with too many past Labour voters, for example, before any turnout step is applied. So the raw answers already contain much of the pollster's judgement about who its respondents are. Truly unweighted answers are available only from Ipsos's current tables and, among historical polls, Qriously's 2019 respondent-level data. Weighting matters for precision as well as bias: unequal weights reduce the effective sample size, the design effect described by Kish (1965), so a weighted sample of 2,000 carries less information than its size suggests.

Raw answers are not identical across pollsters. The S1 base differs. Survation's excludes people at 0 out of 10; Opinium's includes only people who "might vote"; Ipsos's and Find Out Now's include everyone. The historical set is more varied still, because the widest table a pollster printed was not always an all-adults table. The stored detail for each historical S1 figure records what it is: for ComRes/Savanta in 2017 and 2019 it is weighted likely voters; for Savanta in 2024 likely voters before the squeeze; for Opinium in 2019 those who might vote; for Norstat in 2024 the table after its turnout filter but with don't knows included; for Survation in 2015 its forced-choice question; for YouGov in 2015 the constituency question with don't knows and non-voters already excluded. These are the closest available approximations to S1, and for some polls they are already partly through the turnout step. The raw error is therefore partly a property of the mix of pollsters in each year.

Raw answers depend on the question. Three of the six current pollsters ask a question naming the respondent's constituency, which prompts for the parties standing locally; three do not. The two versions can give different answers, as YouGov's own tables show when it prints both. Find Out Now changed its wording during 2026.

Raw answers are not the truth. The raw answers have missed too, and on average by more than the headlines. The claim is not that they are better, but that their errors have been more consistent in direction, which makes them easier to correct. Whether that consistency continues is an empirical question that the next election will test.

Limits

The stage model records what pollsters publish, and pollsters differ in what they publish. For Find Out Now, More in Common and Opinium there is no printed S2, so their turnout and don't-know steps are seen only as a single difference between S1 and S3. For YouGov the whole chain after S1 happens inside a model. For More in Common, part of the don't-know step is a model that is not tabulated.

The England derivation is exact for columns and region sums and approximate for population shares. The approximation has been validated against bounded estimates and is checked against stored figures with a wider tolerance, but it is still an approximation, and it currently applies to one of the six pollsters.

The historical dataset is small: 18 polls with both stages, no more than six in any year, and its S1 stage is defined somewhat differently from poll to poll. The finding that raw errors have a steadier direction than headline errors rests on four elections. It is suggestive and it has been tested out of sample (see Correcting the raw answers: the method behind our UK projection), but it is not established in the way a finding from many elections would be.

Finally, the stage model covers only the six pollsters whose tables we can read automatically. Others publish tables that are harder to parse, behind barriers we do not bypass, or not at all. Their headlines enter the published-poll average but not our reading of the raw answers, and the difference that makes is shown step by step on the home page.

References

Kish, L. (1965). Survey Sampling. New York: Wiley.


  1. For this poll the site's stored GB figures are Survation's published UK-wide headline, because the poll was recorded before the stage tables were loaded. Polls added from their own tables since then store GB as equation (4) gives it. The England figures, which are what the forecast uses, are unaffected.↩

  2. The primary series of the companion paper, which differs by the polls noted above, gives mean headline errors of −2.6 for the Conservatives and +0.7 for Labour, and a 2024 Labour headline error of +4.3.↩

How to cite

Polling Forecast (2026). From answers to headlines: the stage model of a published British poll. Retrieved 20 November 2026 from https://www.pollingforecast.com/research/uk-stage-model. Live figures change as new polls arrive; the date above is when the text was last revised.

You may also like