The Curtice-Firth Methodology of Exit Polls
At 10pm on election day, the polls close and the broadcasters reveal what the exit poll has predicted, estimating how many seats each party will take.
This moment of television is made possible by an innovative statistical method. It also relies on the hard work of polling companies, academics and journalists against the ceaseless march of time.
A model of change
Pre-election polls ask how people intend to vote. The exit poll asks people how they have actually voted, as they leave the polling station. People are asked to fill in a duplicate ballot paper, and place it inside a box.
This survey is then used to estimate seats. In 2017, the broadcasters’ exit poll was commissioned by the BBC, Sky and ITV, and conducted by GfK and Ipsos MORI. In 2019, following a partial purchase of GfK, Ipsos MORI will conduct the exit poll.
Since 2001, the exit poll followed a new methodology, and has a remarkable record of accuracy. This article explains what this method is, and its key ingredients. The method is named after Prof Sir John Curtice and Prof David Firth.
The last election results: The starting point for this method is the last election results. In each constituency, we know (with very little error) how many people have voted for each party. This is our essential resource of voter behaviour.
The previous exit poll: The reason that exit polling is extraordinarily difficult in the United Kingdom is that there is no mandate to count votes by polling stations.
The Curtice-Firth methodology crucially measures change between exit polls. Since (mostly) the same polling stations are used, biases from the choice of polling stations, differential non-response and postal votes are largely cancelled out.
The latest exit poll: The exit poll will use an experienced market and social research company, Ipsos MORI, to conduct the duplicate ballot. The exit poll panel (which polling stations are visited) is carefully planned. High-quality data underpins the entire exercise. Since the polling day occurs during a week, votes are often cast during the early evening.
Electoral change: The concept of the Butler swing is widely accepted, but is not sufficient for measuring change in multi-party contests. In the case of a three-party contest between the Conservatives, Labour and the Liberal Democrats, we need to measure:
- the change in the Liberal Democrat share of the three-party vote;
- the change in the Labour share of the two-party vote.
This can be naturally extended to more parties, such as including the SNP in Scotland.
Statistical model: The survey collects around 200 responses from each of the 140–150 different polling stations. The largest part of the responses comes after 7pm, meaning there is very limited time to analyse. This is then used to forecast 650 electoral contests.
The move from data to forecast requires a model. This is the hardest part: the academic researchers must — with a strict time deadline — use various constituency-level data (such as census variables and the last result) to ‘explain’ patterns of electoral change. One example might be larger changes in the Labour vote share between urban and rural constituencies.
Probability, not certainty: The exit poll analysis is then used to estimate what the vote share in each constituency will be, and what is the probability of each party winning that seat.
In an example seat where Labour are ahead of the Conservatives, the win probability for Labour is 70%, and 30% for the Conservatives. In the final forecast, this example seat counts for 0.7 towards Labour’s total, and 0.3 for the Conservatives.
These values are then summed, giving us the seat estimates we see on our televisions at 10pm. The total estimates are subject to uncertainty. In 2017, the credible interval for Conservatives seats was 297–330 seats, with a central estimate of 314. Since this crossed the boundary of an overall majority, the summary statement was ‘Conservatives largest party’, not ‘hung parliament’.
For 2019, the exit poll’s analytical team is: Curtice, Fisher, Kuha, Mellon, Ford, English, Jennings, and Prosser.
No statistical method can offer perfection — but the Curtice-Firth methodology of exit polling overcomes major issues in how the country counts votes.