David raises a number of important concerns about the way conventional methods allocate undecideds. In addition to David's suggestion that we consider more accurate ways to assess voting intentions, we might also consider how the campaigns themselves are likely to influence the accuracy of pre-election polls.

The final 72 hours leading up to the close of polls on election day will feature a targeted barrage of phone calls, door-to-door contacts, and emails meant to rally each party's likely supporters to voting booths in key battleground states. Depending on the outcome of what campaign professionals call the "ground game," survey predictions of state-level presidential outcomes could understate a candidate's likelihood of winning targeted states by failing to account for the way individual voters are targeted by the campaigns for special "get out the vote" (GOTV) attention.

The "ground game" is the systematic effort made by a campaign to deliver its likely supporters to the polls on election day. Before 2004, these efforts were coordinated by the presidential campaigns and the national party committees, but the workings of the ground game were largely left up to state- and county-level "get out the vote" (GOTV) operations staffed by volunteers and run by local party officials. Beginning with a test run in the 2002 congressional elections and culminating in a nationally-organized GOTV operation unprecedented in modern American history, the Republicans developed a highly organized get-out-the-vote-operation that allowed individual voters to be targeted directly by the national campaign. Likely Bush supporters living in battleground states were given extra attention to ensure that they would show up and cast a ballot on election day. Republican operatives claimed after the 2004 election that their ground game was decisive in re-electing President Bush.

Fast forward four years. The ground game will be ramping up on both sides in little more than a week. By all indications, the effort to mobilize Obama supporters seems likely to out-hustle the effort being made to target McCain supporters. Will it affect the accuracy of pre-election polls?

Here is how the ground game could potentially influence how polls predict likely state-level winners in the presidential contest. By this time in the fall campaign season, most polling outfits use a variety of methods to filter out survey respondents who are unlikely to vote on election day. This is done for a good reason: If nonvoters were left in these surveys, their preferences--if different from the mix among people who will actually turn up at the polls--would bias the results of pre-election surveys gauging likely levels of candidate support come election day.

Details on how different organizations filter their survey results to identify likely voters or registered voters are available elsewhere (two especially good sources are Mark Blumenthal's "Mystery Pollster" posts from the 2004 election and Frank Newport's discussion of how Gallup screens for likely voters in 2008). The important point is that none of the available methods takes account of the ground game. The reason is that pollsters build their likely voter filters using information provided by survey respondents about such things as their registration status, past voting history, and degree of interest in the presidential campaign. None of these is a good indicator of whether the survey respondent is slated to be contacted by one of the party's GOTV efforts in the final hours of the campaign.

Since each campaign's ground game is aimed at those people who are likely supporters but unpredictable voters, the people most likely to be targeted for GOTV effort are those most likely to be filtered out by traditional screening methods. Registered partisans who have an uneven voting history and limited interest in the campaign are both most likely to be dropped from pre-election polls by conventional likely voter screens and most likely to be targeted as part of a campaign's ground game. If they live in battleground states.

But if the state-level results from 2004 are any indication, the degree of pre-election polling error generated by the ground game seems unlikely to be large.

According to Daron Shaw's account of his work as electoral college strategist for the Bush-Cheney campaign, the key battlegrounds being targeted by the Republicans in 2004 included Maine, Washington, Oregon, Michigan, Minnesota, West Virginia, Florida, Pennsylvania, New Hampshire, Iowa, New Mexico, Ohio, Wisconsin, and Nevada.

For each of these states, I tracked down average pre-election poll results calculated by Real Clear Politics from polls taken in the final week before election day. Subtracting the expected Bush percentage of the vote from the expected Kerry percentage yields the predicted result for each state. I did the same calculation using the actual state-level vote totals reported by the Federal Election Commission.

Figure 1 shows that the final-week pre-election polls in 2004 were predicting Bush losses in seven of the 14 battleground states. In this figure, states are ordered from left to right according to the predicted Bush margin of victory, so the seven states predicted to go to Kerry are all toward the right side of the graph. However, Bush ended up losing eight of these battleground states: all seven predicted by the pre-election polls plus Wisconsin, which had appeared to be narrowly leaning toward Bush before election day.

In these battleground contests, the differences between pre-election predictions and the actual outcomes were sometimes far from trivial. Take Florida, for example. Averaging across the final week's pre-election polls suggested that Bush would win by less than a percentage point, 48.2% to Kerry's 47.6%. But the actual result was that Bush took 52% of the vote to Kerry's 47%.

It is tempting to conclude that the Republican ground game was responsible for this unexpected Florida margin favoring Bush. If so, and if the Republican GOTV efforts were directed in even measure toward likely supporters in each of these battlegrounds, then we should see that the differences between predicted and actual outcomes in the battlegrounds consistently favored Bush. These differences are hard to track in Figure 1, so Figure 2 arranges states from largest to smallest according to the size of gaps between predicted and actual outcomes in Figure 1.

Figure 2 helps us to see that although the pre-election polls were sometimes off the mark, there was no consistent error that understated the actual Bush vote. Seven states swung further toward Kerry on election day than predicted in the final wave of pre-election polls, six swung more toward Bush than expected, and Ohio's outcome was predicted spot-on.

If the Republican ground game made a last-minute difference in the battleground states, it does not show up consistently in these data. Bush did four points better in West Virginia and Florida than was predicted in the final wave of pre-election polls, but he did three points worse than predicted in Washington and four points worse in Nevada. Figure 2 shows no consistent pattern of unexpected, election-day movement toward Bush that would be produced by a successful ground game.

Are current estimates about state-level contests between McCain and Obama potentially off the mark? That depends on a number of factors, including how late-deciding voters end up splitting between the candidates and whether younger voters turn out at higher rates than is typical in presidential elections, as seems likely to happen. If the level of organizational complexity in the Democrat's GOTV operation is any indication of its effectiveness, the ground game could be poised to give a widespread, last-minute boost to Obama's margins, a boost that will not have been anticipated in the pre-election polls.

But the evidence from 2004 suggests otherwise. A highly-oiled and fine-tuned Republican GOTV operation did not appear to yield unexpected election-day surges favoring George Bush.

And this year, unlike 2004, the Republican machine won't be the only game in town. The Democratic operation will be counter-mobilizing in full force, and these countervailing efforts could even out any unanticipated effect of the competing ground games. Unfortunately, since this work is going on in secret, there is no way to tell until election day which way the ground games will break.