What election polls really tell us

Political polling has long been an interest of mine. It's an interesting practice to say the least, as we take a relatively small number of opinions and try to predict effects for a whole state or for the entire country. Prediction gets particularly dicey when the polls suggest a close race, as they did leading up to the 2004 Presidential Election. While some polls suggested a John Kerry lead late in that election, Bush's final margin on a national level was less than 3%, which was within the margin of error for most polls. The next 48 hours will provide an interesting comparison to see how the poll numbers actually reflect the final outcome. It might seem as though McCain's camp's claim that things are tightening are silly given some of the predictions of Obama victory. Despite efforts to build likely voter models that attempt to determine the margin among those who think they will vote, only the ultimate election outcome will determine who is President. I know that sounds so simple that it's practically stupid, but the huge emphasis placed on polling by the media means that it's worth remembering that telephone polls and voting are two different measures of public opinion.

Things to watch with polls:
Do they include cell phone only households? An increasing number of households do not have land line telephones. Three times more households have no land line phone in 2008 than in 2004. These voters tend to be younger and more likely to support Barak Obama. Many current surveys do not include cell phone only respondents, which could be underestimating support for Obama. I think this effect will be somewhat offset by the fact that cell phone only voters tend to be younger and in the past have not turned out to vote in large numbers.

Percentage of sample from each party. It would seem on the surface that a political poll should include an equal number of Republicans and Democrats to be fair right? That would be true if you wanted the poll to be fair and not to reflect reality. Reality is that nationwide there are more registered Democrats than Republicans. Nonetheless, some polls are using voter turnout numbers from the 2004 election, when Republicans represented a larger number of the electorate than Democrats. One that uses current data on voter party identification is Rasmussen Reports. Polls that use old party identification data will also underestimate support for Obama.

Who will show up to vote? You can have all the polls in the world, but unless you can predict who will show up tomorrow, you're dead in the water. Despite this obvious statement, many polls, including Gallup, use a likely voter model that only counts responses for those who voted in the 2004 election. Gallup has an alternative model that bases likely voter status only on if the person intends to vote this year. The difference between the two approaches is a 2 point advantage to Obama on average.

We'll look at the polls, particularly at the state level, in the next week to see how close they were.