Some respondents choose a party immediately. Others say they do not know, refuse to answer or say they will not vote. Those answers create a substantial part of the work between an interview and the headline voting-intention figure.
Two polls with similar initial answers can produce different headlines because they treat these respondents differently. The differences need to be read in the methodology, rather than inferred from the final percentages alone.
Removing undecided answers changes the denominator
In an illustrative sample of 1,000, 250 people choose Party A, 200 Party B, 350 other parties and 200 give no preference. Removing the final group gives Party A 250 out of 800, or 31.25%, rather than 25% of the whole sample.
That is a share among people who named a party. It does not establish how the other 200 would vote. Nor does it show that all 800 decided respondents will eventually cast a ballot.
The distinction appears throughout polling tables. An all-adults question, a decided-voter calculation and a modelled headline can each use a different base.
Follow-up questions and modelling
A follow-up, sometimes called a squeeze question, asks an initially undecided respondent for a preference under a more pressing formulation. It supplies additional answers, but some people remain undecided even after the follow-up.
More in Common’s January 2026 MRP methodology describes both a follow-up question and modelling of remaining undecided voters using information about similar respondents. That method estimates behaviour which was not directly supplied as a firm choice.
Reallocating undecideds is not automatically more accurate than excluding them. Both approaches rest on assumptions. The difficulty grows when undecided people differ politically from the people who have already chosen, or when their indecision reflects a real change of allegiance.
Who is likely to vote?
A respondent’s enthusiasm for answering political surveys need not match their enthusiasm for voting. Pollsters can use stated likelihood, past behaviour, demographic information or combinations of these when estimating turnout.
A turnout weight also changes the balance between parties when their supporters have different probabilities of voting. In a purely illustrative calculation, 300 supporters of Party A with an average voting probability of 0.8 contribute 240 expected votes; 300 supporters of Party B with probability 0.6 contribute 180. Equal starting support becomes unequal expected voting support.
Those probabilities are assumptions for the example, not measured turnout rates for real parties. The model needs evidence for whatever probabilities it uses. Actual election turnout is a separate quantity, calculated from the election returns.
Why the assumptions matter after an election
YouGov’s review of its 2024 performance found that its sample included too many voters and too few abstainers. The review discussed how different turnout modelling would have affected its estimates. That is useful evidence about a particular election, rather than a guarantee that one method will always work best.
When comparing releases, check whether turnout and undecided treatment changed between them. A jump after a method change cannot safely be read as voters moving by the same amount.
These decisions help explain persistent differences between pollsters. They also belong in an assessment of polling accuracy, where the question is how well the complete process performed, not whether the original sample was large.
Sources and further reading
Reviewed 7 October 2026. Worked examples are illustrative unless an election or poll is identified.