General Election 2001

Opinion polls measure the electorate's intentions in votes, not in seats. We can -- if nothing goes wrong -- measure voting intention percentages directly, and would hope to be accurate within our margins of error. But projecting the number of seats that a given share of the votes would give, although it produces a better headline, involves much greater uncertainties, and we rarely have the information we need to produce such figures with anything approaching precision.

Because a small shift in votes can cause a large shift in seats, the margins of error sound much larger when translated into seats, and even very similar polls can produce substantially different seat projections. Consider the figures from our poll in tomorrow's Times [Times 2001 Campaign Polls]. In votes, this has the Conservatives on 30%, Labour on 54%. Incorporating the margin of error, this means Conservative 30+/-3%, Labour 54+/-3%. But translating this into seats assuming uniform swing, it places the overall majority anywhere between 231 and 363. (In fact the difference is between Labour winning 68% and 78% of the seats, so in percentage terms it is only a margin of error of +/-5%, but it sounds like a lot more than that.)

And that is assuming that the swing really is uniform and national. That is the conventional method for producing seat projections, but it is almost certainly an inaccurate simplification; however, it has the merit of being a straightforward and defensible basis for analysis. Vote shares on their own are somewhat abstract; it helps to explain them to express the figures in the terms that if everybody in the country behaved in the same way, the outcome in seats would be this.

Then you can go into the likely exceptions. There may be regional variations. Rural and urban seats may behave differently. The target marginal seats, where the parties concentrate their campaigning efforts, may swing differently from the neglected safe seats. There may be differential turnout. There may be tactical voting in some constituencies -- or, rather, there may be more or may be less than at the last election.

But most of these factors are difficult to measure if you prefer to attempt a more sophisticated projection model. No single opinion poll can give you very much useful information about how far voting will deviate from uniform swing, although it may give some indication of differential turnout; even an aggregation of a number of polls gives only broad indications, and while it may detect systematic regional variations, it can't deal with tactical voting, let alone the peculiarities of individual constituencies. (In fact, in recent weeks the turnout data from our polls have suggested little party effect -- see Low Turnout -- Who Loses?, and although the regional variations are substantial the effects in seats mostly cancel each other out; but these patterns may change.)

It is, of course, possible to build an alternative projection model, building in various assumptions about all of these factors. ICM, bravely, have done so, based on the evidence from their polls. (The ICM model suggests that if the vote shares were the same as in 1997, Labour's majority would be cut by 60, and that the bias in the electoral system has switched from favouring Labour to favouring the Tories, so that the Tories would be the bigger party if votes were equally split.)

We don't feel that the polling evidence strong enough at this stage to make any heroic assumptions. Nor would it be wise to rely on the evidence of voting patterns in the European and local elections, or for the Scottish Parliament and Welsh Assembly, when turnouts were much lower than is likely at even this general election and the party political issues different. So we shall stick to a uniform swing projection, to which we may then add caveats. But it is a "what if" projection, NOT a prediction.