Political observers have become more sophisticated in their reading of polls in recent years. They know enough now not to read too much into summer polls. Poll leaders in June are not any more likely to win than their opponents who trail them in the summer polls. By the time of the conventions, though, polls become more meaningful. Since 1948, 12 of the 15 candidates leading in the Gallup poll after the conventions have gone on to win the national popular vote, and by late September the poll leaders have a record of 14 wins and one popular vote loss (Tom Dewey in 1948).
There is more to learn about being sophisticated poll readers, however, than knowing when to take polls seriously. If we compare polls at points in the campaign to the eventual vote, it is clear that a poll lead among registered voters at some point in a campaign is not the same thing as the vote lead on Election Day. Poll leads are, however, related to vote leads in a systematic way, more closely related later in the campaign than earlier, but related throughout the campaign. The question is how are they related? How can we draw the maximum information out of the polls rather than reading them as though the percentages that the polls report are the best indications we have of the vote percentages that the candidates will receive?
From examining the polls over the years, I have two recommendations. The first corrects for a systematic difference between poll respondents and actual voters. Pollsters have been trying to figure out for many years now how to best measure the vote intentions of people who will actually vote. They have tried to determine who are likely voters. This is not so easy—especially weeks and months before the election. Pollsters ask their respondents their voting history, their interest in the election, their intention to vote, and so forth; on this basis, they try to determine who among their respondents are likely to vote. But these guesses about likely and unlikely voters are highly error prone. Despite the difficultly of weeding out nonvoters from the ranks of poll respondents, pollsters admirably continue to try to do so because they know that including the preferences of nonvoters can distort the measure of what real voters are thinking.
I would like to suggest an alternative. Rather than trying to correct the survey data at the individual level, let’s make the correction at the aggregate level where we have better data to determine if the correction is worthwhile. In particular, we know from decades of research on voter turnout that registered Republicans are more likely to vote than registered Democrats. In every presidential election since 1952, according to National Election Study surveys, registered Republicans were more likely than registered Democrats to report having voted. The average difference was over four percentage points. The typical Republican is more likely than the typical Democrat to have the social and economic characteristics associated with voting. Republicans tend to be better educated and more affluent. We also know that partisans tend to vote for their party’s candidate. Party loyalty rates are quite high in presidential voting and have increased in recent years. In 2004, 90 percent of Democrats voted for John Kerry and 93 percent of Republicans voted for George W. Bush.
The fact that partisans are likely to vote for their party’s candidate and that registered Republicans are more likely to vote than registered Democrats implies that polls of registered voters underreport the vote intentions for Republican candidates and over-report the vote intentions for Democratic candidates. Now we are not talking big differences, there are plenty of wealthy and educated Democrats and less educated and poor Republicans, but the differences are probably big enough to take into account, especially in close elections.
Is there a pro-Democratic tilt in polls of registered voters? To determine whether there has been, I examined the in-party candidate’s level of two-party support in Gallup polls of registered voters conducted at ten points in the fifteen presidential campaigns from 1948 to 2004. The ten points in the campaign were June, at the fourth of July, later in July, just before the first convention, just after the second convention, around Labor Day, in mid October, in early November, and the final pre-election poll. In a few of the early years, because of the infrequency of polling, the same poll was occasionally used in two points of a campaign. The correlation of these polls with the November vote was then examined in two ways: first without any adjustment and then again after adjusting the poll by adding one point if the in-party candidate was a Republican and subtracting one-point if the in-party candidate was a Democrat. The results were conclusive. At each of the ten points in the campaign, the correlation between the adjusted poll and the eventual vote was stronger than between the unadjusted poll and the vote. Ten for ten. The differences are not large, but they are absolutely consistent. My advice: if you want to know where the voters really stand at any point in the 2008 election, ignore likely voter polls and, in polls of registered voters, subtract one point from Obama’s two-party percentage and add one point to McCain’s.
My second recommendation is that poll leads should be discounted by more than half early on in the campaign and by decreasing amounts as you approach the election. Even in October and November, the history of the polls indicates that the front-running candidate’s poll lead is likely to be trimmed by about a third at the ballot box. Before the conventions, the historical discount rate for polls is about 55 percent. That is, you get the smallest error between the pre-convention polls (after making the above party correction) and the eventual vote by discounting the frontrunner’s poll lead by 55 percent. A candidate standing with 56 percent of two-party support in pre-convention polls more realistically has about 52.7 percent (56 – (6 x .55) = 52.7) of the two-party vote among those who will actually cast a vote.
So where do the candidates stand at this point? Averaging the first ten Gallup tracking polls conducted in August, Obama leads McCain by 46.0 to 43.1. This gives Obama 51.6 percent of the two-party trial-heat support to 48.4 percent for McCain. Drop one point from Obama and add one to McCain (the actual versus registered voter correction) and it is 50.6 percent Obama to 49.4 percent McCain. Using the historical association between pre-convention polls and the eventual vote, the six-tenths of a point lead should be expected to shrink by about half.
Of course, even after adjusting for partisanship and discounting the frontrunner’s lead, the polls can only tell us so much. Even corrected polls are imperfect, and there are always a few late developments and last minute deciders. The adjusted polls as late as the week before the election have an average error of plus or minus one point and in some elections have been off by more than two points. Before the conventions, the average error is larger—about two and half points. So, our best reading of the voters at this point is that they are divided about 50.3 percent for Obama to 49.7 percent for McCain—give or take about two and a half points either way. Unless the conventions, the debates, and later developments tilt predominantly one way, the odds are that we may be spending much of November pondering whatever becomes the 2008 equivalent of hanging chads.
Jim Campbell is a professor and chair of the Department of Political Science at the University at Buffalo, SUNY. He has published four books, nearly fifty articles in scholarly journals, and numerous book chapters. His most recent bookis The American Campaign, Second Edition: U.S. Presidential Campaigns and the National Vote (Texas A&M University Press, 2008).