Election 2010: It ain’t just that swing

At the weekend, Comment is Free ran this piece from me which was a follow up to my poll reading tips post:

Election 2010: It ain’t just that swing
Cif reader glassfet asked why ‘uniform national swing’ probably won’t help to predict the result. Here’s why.

The number of seats each party wins is only very loosely related to how many votes it gets across the country. Usually more votes means more seats and vice versa, but not always. In 1997, for example, the Liberal Democrat vote share fell slightly but the number of seats won leapt upwards. That is because each seat’s result is determined by just the votes that are cast in that seat.

So if you think you know what the overall shares of the vote will be, how can you work out what the numbers of seats will be for each party?

The simplest answer is “uniform national swing” (UNS). Suppose you think Party A’s support will go up by 10% compared with the last election, Party B’s down by 8% and Party C’s down by 2%. You then just work through every individual constituency, making those same changes to the vote share in each and see who the winner is.

Of course, that’s never what actually happens. There are all sorts of factors which make the actual changes vary from seat to seat: a big housing redevelopment in one, a popular candidate in another, a hospital closure in a third and so on.

As a result, even UNS predictions based on the correct vote shares are far from perfect. In the last three general elections, UNS predicted the Conservatives would get 42 seats more than they did (1997), 15 more (2001) and 13 less (2005). For Labour UNS predicted 23 too few (1997), 10 too few (2001) and 14 too many (2005). For the Liberal Democrats it was 18 too few (1997), 5 too few (2001) and spot on (2005).

For a simple predictive tool, those errors are not too bad – if the election is not that close. The 42 seat error on the Tory prediction did not really matter in the landslide conditions of 1997; in 2010 a similar error could mean the wrong winner is predicted.

Despite these errors, UNS predictions are usually quoted to the nearest seat as if they are really that accurate. They are not. There are other, more sophisticated swing models, but they are either so far untested in a general election or have also had errors in their past predictions.

So if the election is looking close, we won’t know the result until after the polls close, no matter how precise the predictions may sound.

• This article was written at glassfet’s suggestion