ELECTIONATE SEPTEMBER 26, 2012
Campaign 2012 has reached the stage where partisans have moved from attacking their opponents to attacking the polls. And while there's been criticism from all sides, the current state of the polls explains why the most vociferous denunciations are coming from the right. One dismayed conservative has even set up a website, unskewedpolls.com, where poll results are re-calibrated in ways that supposedly strips them of their bias.
One common complaint in the current conservative fusillade is that many 2012 polls are using the “2008 turnout model.” Karl Rove, for instance, alleges that the pollsters are weighting their surveys to reflect the partisan and racial composition of the 2008 electorate–when Democrats outnumbered their Republican counterparts by 7 points on election day. Conservative critics think the GOP’s enthusiasm to oust President Obama means that differential will be a lot smaller this year. That might be valid, but the implication that polls are rigged to reflect the 2008 electorate is outright misleading: most of this year’s polls don’t use “the 2008 turnout model.”
In fact, the “2008 turnout model” critique is so far off base that responding to it simply entails explaining how polls work.
Most pollsters don’t weight their polls to match a preconceived electorate. Instead, they take a demographically representative sample based on actual figures from the US census and then let respondents speak for themselves about whether they’re voting for Obama or Romney. For illustrative purposes, consider the Bloomberg/Selzer poll. They started by taking a sample of all American adults, weighted to match the demographics of all adults in the US census, like, race, education, and marital status. To produce a likely voter sample, they then would have excluded adults who weren’t registered to vote and then asked a series of questions to help determine who was likely to vote.
Ultimately, Selzer’s sample found Obama leading by 6 points, by 49-43. Whatever you think of the outcome, it wasn’t the result of Selzer imposing her assumptions upon the sample; she let her sample speak for itself. Did she take a good sample? We’ll find out on Election Day. But if she’s wrong, it won’t be because she used the “2008 turnout model.” In fact, this particular poll had a relatively innocuous Democratic advantage of 6 or 2 points among likely voters, depending on whether you include leaners. But even if Selzer had shown something “crazy” like R+5 for D+15, remember that the party-ID of the sample wouldn’t have been determined by Selzer’s premonitions.
This isn’t to say that the 2008 election doesn’t factor into the polls at all. Some likely-voter models use past voter participation to help determine whether someone will vote in 2012. These polls might, conceivably, under-represent the number Obama ’08-Couch ’12 voters.
To be sure, there actually are some polls that might use something like a “turnout model.” For the most part, these are cheap, automated surveys. Why? Many of them terminate interviews early, sometimes as early as learning that a voter doesn’t intend to vote. That makes interviews cheaper, but it prevents them from weighting their sample to figures in the U.S. Census. But at any rate, this type of poll doesn’t include the top, live-interview pollsters—it’s second-tier stuff, like Gravis Marketing, who actually says that its poll is weighted to the “anticipated voting proportions for the 2012 General Election.” Unfortunately, the poll-bashers haven’t made that distinction when publicly insisting that the results are skewed by bias.
We don’t need pollsters to tell us that Obama would lose if the electorate looks like 2010, nor do we need them to tell us that Obama would win if the electorate looked like 2008. What we need—and what we have—are pollsters with methods that allow us to get a decent grasp on what’s going to happen on Election Day. Pollsters are not sooth-sayers who correctly guess the composition of the electorate every four years; they take demographically representative samples of adults and let the sample speak for itself. That’s how polls using the same methodology managed to show Bush winning in 2004, Obama winning big in 2008, and a GOP takeover in 2010. It’s how Ann Selzer managed to show Obama winning the Iowa Caucus’ in 2008, even though there wasn’t any comparable Caucus to mirror as a “turnout model.” None was necessary.