At 10pm on May 5th the BBC and ITV predicted that Labour would win the election with a majority of 66. A few hours later that was exactly the result that transpired. Labour had a majority of 67 seats, a majority that will fall to exactly 66 assuming the Conservatives eventually win the delayed election in the safe seat of Staffordshire South.
This accurate prediction, based on an exit poll jointly commissioned by BBC and ITV News, was the work of a small team led by Professor John Curtice (Government, Strathclyde University) and Professor David Firth (Statistics, University of Warwick; currently an ESRC Professorial Fellow). Also closely involved were Professor Neil Shephard (Economics, Oxford), Professor Colin Rallings (Politics, University of Plymouth), Dr Stephen Fisher (Sociology, Oxford) and veteran psephologist Clive Payne (formerly Oxford, now retired).
The exit poll was designed with the aim of estimating patterns of vote change across the country, including in both marginal and non-marginal seats. A total of 120 polling stations were covered on election day, and around 20,000 voters interviewed. Detailed statistical modelling of patterns found in the exit-poll data then led to an estimate for each party of their chances of winning each House of Commons seat. The final forecast of a 66 seat Labour majority was the result of adding up all those seat-by-seat predicted chances (i.e., probabilities) for each party.
Fieldwork for the exit poll was done jointly by two commercial polling organisations, NOP and MORI, who then supplied the data to the analysis team.
The pattern of electoral change indicated by the exit-poll returns on the day was far from obvious: broadly the picture was of a 3% swing from Labour to the Conservatives, but with Liberal Democrat performance highly variable from one place to another and overall less strong where it was in competition with the Conservatives. Conventional "swingometer"-style calculations of the consequences of a 3% swing would have implied a Labour majority of more than 100 seats. But the use of careful statistical modelling, to allow both for the substantial systematic variation in swings between different types of seats as well as the likely level of random variation, meant that the BBC/ITV team was able to predict more accurately the number of seats that Labour would lose on election night.
Getting the majority exactly right was, of course, largely a matter of luck. The same methodology used in 2001 led the BBC to a forecast within 10 seats of the eventual majority (as was ITV's), based on a rather smaller exit poll. Any exit-poll based prediction within 20 seats of the eventual majority should be viewed as remarkably successful, and by that criterion the statistical methods as used for the last two General Elections have now clearly proved themselves.
The 3-party "electoral triangle" below shows just how much variation there was in the pattern of change: the green points are 2005 shares of the 3-party vote for each constituency, the black tails point to the corresponding 2001 shares.
Copyright © 2005 David Firth, all rights reserved