How do opinion polls and seat projections work?

An opinion poll estimates what a larger population thinks by interviewing a sample of people.

A seat projection goes further: it tries to estimate how voting support might translate into winners across the UK's 650 Parliamentary constituencies.

They are related — but they are not the same thing.

How can a poll represent millions of voters?

Polling organisations interview a sample rather than the whole electorate.

The sample is then selected and/or weighted so that it better reflects the population the poll is intended to represent.

Factors can include:

• age;
• sex;
• region;
• past vote;
• likelihood of voting;
• other demographic or political characteristics.

Different polling companies use different methods.

That is one reason two reputable polls carried out at similar times can produce somewhat different results.

How large is a normal poll?

Many standard British voting-intention polls use samples of around 1,000 to 2,000 people.

A sample does not have to contain a fixed percentage of the whole population to be useful.

What matters is the sample size, how respondents were selected, how the data were weighted and how well the sample represents the population being estimated.

Are UK polls actually UK-wide?

Not always.

Many voting-intention polls cover Great Britain — England, Scotland and Wales — and exclude Northern Ireland because Northern Ireland has a different party system.

That is perfectly valid if the poll is clearly labelled.

But users should check the population represented before calling a Great Britain poll a poll of the entire United Kingdom.

Are polls exact?

No.

All polls contain uncertainty.

Possible sources include:

• sampling variation;
• who agrees to participate;
• weighting choices;
• turnout assumptions;
• treatment of undecided voters;
• question wording;
• late changes in opinion.

The British Polling Council requires member polling organisations to publish extensive methodological information so users can assess poll evidence.

Treat a poll as an estimate, not as a precise count of future votes.

What does a change of one point mean?

Not necessarily very much by itself.

Suppose Party A is reported at:

31% in one poll

and:

32% in the next.

That does not automatically prove the party has gained support.

The difference could reflect normal polling variation.

It is usually more informative to consider:

• polling averages;
• repeated movement across several polls;
• methodology;
• fieldwork dates;
• changes larger than ordinary polling noise.

How do we turn national polls into seats?

This is the difficult part.

Because FPTP works constituency by constituency, analysts need a model.

A very simple method might use uniform national swing:

1. estimate the national movement between parties;
2. apply roughly the same swing to each constituency;
3. identify which seats would change winner.

That is easy to understand, but real electoral change is rarely perfectly uniform.

What is MRP?

MRP stands for:

multilevel regression and post-stratification.

In simple terms, an MRP model:

1. uses a large survey to estimate how different types of people are likely to vote;
2. considers characteristics such as age, education, location and previous vote;
3. applies those estimated relationships to the known population profile of individual constituencies;
4. produces constituency-level vote estimates or winning probabilities.

MRP can therefore model local differences that a simple uniform-swing calculation misses.

Does MRP tell us exactly who will win every seat?

No.

It is still a model.

Close constituencies can be highly uncertain.

Different MRP models can produce different seat projections because they may use:

• different polling data;
• different variables;
• different assumptions;
• different turnout models;
• different treatment of smaller parties.

A projected seat total should therefore be read as an estimate with uncertainty, not a guaranteed result.

Why can a small polling change create a large seat change?

Because many seats may be close to the point at which they change winner.

Imagine 50 constituencies where the previous winner is only two or three points ahead.

A fairly small movement in votes could move many of those seats simultaneously.

This is why the relationship between polling and seats can look non-linear.

One extra point of national vote share does not always equal the same number of extra MPs.

What should I check when reading a poll?

Useful questions include:

1. Who conducted it?
2. Who paid for it?
3. When was the fieldwork carried out?
4. Who was interviewed?
5. How large was the sample?
6. What population does it represent?
7. How were respondents weighted?
8. What exact question was asked?
9. Is the apparent change larger than normal polling uncertainty?

Related explainers