For several weeks, polls of “likely voters” have told us conflicting stories: National surveys see a presidential race that is neck-and-neck, or possibly Mitt Romney leading by a nose, while state-level polling credits President Obama with a consistent edge, especially in the Electoral College.
There are several possible explanations for the discrepancy. In a couple of days, we’ll know which one is right.
One possibility, proffered by strident partisans on both sides, is that polls showing the other team ahead are exercises in propaganda rather than public opinion research. There is clearly at least a small grain of truth to this assertion. Some pollsters have clear party affiliations, while others are independent organizations that nonetheless have a relatively consistent tendency to lean in one direction or another. Analysts have a term for this: “house bias.”
Polls taken internally by the two campaigns and the political parties, which are occasionally shared with the press, have likewise tended to diverge – at least in the case of those that the campaigns have decided to share. Each side presumably wants to convince its supporters that it is ahead in order to keep them motivated, but not to make supporters believe their man is ahead by so much that complacency sets in. Trash-talking the other side’s prospects also might serve to dishearten the opposition.
Another possibility is that the national and state polls are not as inconsistent as they appear. It is certainly possible that either Romney or Obama might win the popular vote but lose in the Electoral College, as Al Gore did in 2000. This would be most likely to happen if Romney racks up a big vote in deep-red states like Texas, while some Obama voters in safe states like New York and California take the day off. This scenario also implies that Romney’s overall edge in popularity, at least among citizens who take the trouble to actually vote, will not hold up in the tightly contested swing states, and especially those that are part of Obama’s Midwestern “firewall” of Ohio, Wisconsin and Iowa.
A third possibility is that the polls are consistently, unintentionally wrong – either overstating Romney’s strength nationally, or understating it in key state-level surveys. This is what statistician-journalist Nate Silver, who writes the FiveThirtyEight blog on the New York Times website, calls “systematic bias.” And it is what I believe is the most likely explanation for the confusing poll numbers.
I think the basic problem is that most polls try to treat all respondents equally, as much as possible. But some people are more likely than others to share what they are going to do, and some people are more likely than others to actually go ahead and do it.
Let’s start with the basics. If you want to gauge the likely outcome of an upcoming election, you don’t survey people who can’t vote, such as children and foreigners. The biggest possible survey pool is registered voters. Pollsters of all persuasions acknowledge, however, that surveying all registered voters creates an overstated Democratic sample, because registered Democrats tend to vote at lower rates than registered Republicans.
So the pollsters ask their survey subjects whether they are “likely” to vote. As Election Day draws nearer, pollsters and analysts like Silver and the website RealClearPolitics.com pay less and less attention to surveys of registered voters, and greater attention to surveys of likely voters. Pollsters use varying methods to determine which respondents are genuinely “likely” voters, but my understanding is that most pollsters, having determined which voters are “likely,” count them equally. (Some may try to make further statistical adjustments.)
And here is where one potential problem arises. Consider a pool of 200 people: 100 who back Romney and 100 who back Obama. In this pool, hypothetically, personal circumstances such as work schedules and child care, and varying levels of enthusiasm for their candidate, are such that 90 of the 100 Romney voters will actually cast ballots, while only 80 of the 100 Obama voters will do so. Even at the 80 percent level, the Obama backers are “likely” voters. A survey that counts them equally with the Romney backers will thus overstate their strength, finding 50 percent support for the president when, in fact, he will receive only 47 percent (80 out of 170) of the vote.
Now, I simply assumed my way to this result. How can I know, in advance, that 90 of the 100 Romney backers will actually vote, while only 80 of the 100 Obama supporters will do so? I can’t, and neither can the pollsters. The problem is not that the pollsters are trying to skew their samples; it is that the samples may be skewed already, and the pollsters have only limited, imperfect means to detect and correct that distortion.
If the state-level samples are skewed this way, especially the crucial polls in the swing states, are the national polls similarly affected? If the answer were yes, the differences between the two sets of polls would remain unexplained. But I suspect the answer is no.
A socially conservative Republican voter in Texas, who probably did not start this election season as a Romney backer, may not be all that motivated to vote for him. Romney is going to carry Texas anyway, so the likely Romney voter in Texas may not be quite so likely to vote. But a socially conservative voter in Ohio, seeing a very close race in that state and having to choose between Romney and the even less socially conservative Obama, is probably quite a bit more likely to vote than his Texas counterpart. I think this, but I can’t prove it, at least not until after tomorrow’s results arrive.
Romney is drawing his strongest support from demographics that tend to vote very reliably, including whites (where Obama’s support is hovering near 40 percent), older voters, Jews and non-Hispanic Catholics. Obama is running strongly with college-educated voters, but they are a minority overall and a smaller minority in most swing states. His other key groups include African-Americans, young voters and Hispanics, all of which typically have low turnout rates. Obama benefited in 2008 from much better than typical turnout of those groups. The ultimate question of this election is whether he can repeat that performance.
I don’t think so or, more accurately, I don’t think he can do so to the same extent. I suspect, but again cannot demonstrate, that Obama’s get-out-the-vote efforts will make a difference; I just don’t think it will be as big a difference as four years ago.
As a result, I think Obama is highly likely to lose Florida and Virginia, which are rated as close by most analysts, and I have concluded that he will also probably lose Ohio, which is not what the polls have been showing. I think he is more likely than not to lose Wisconsin, despite polls showing him winning there. Wisconsin has voted Republican in statewide races four times in a row since 2010, and I don’t see why the result will be different this week.
Obama could also lose Pennsylvania, especially if the aftermath of Hurricane Sandy holds down turnout in key areas in the southeast part of the state, which is his stronghold. More likely Obama will pull out a narrow victory there, but if he loses Pennsylvania, the election is effectively over. New Hampshire and Iowa could probably go either way. A late GOP advertising effort in Minnesota probably won’t be enough to flip that state to the Republicans.
I am backing Romney in this election, so it is possible that my own personal preferences are biasing my conclusions. I have tried not to let this happen, but I am human, so we have to wait until tomorrow night (at least) to see whether I succeeded.
One final note: I mentioned the private polls that campaigns take and sometimes share. There is a lot of information they don’t share, and I like to watch what the candidates actually do rather than what they say. Obama, until being caught up in post-Sandy recovery last week, looked like a candidate who believed he was on the road to defeat. Romney has consistently campaigned like someone who sees victory coming. I believe both sides have reached similar conclusions to mine, with Republicans already trying to position themselves for jobs in a Romney administration.
The bottom line is that pollsters can only talk to registered voters or likely voters, while elections are decided by actual voters. In most surveys, the determination of a likely voter is probably accurate enough not to matter, especially when both sides are roughly equally likely to vote. But this does not look like that sort of election to me, and the systematic bias that Silver has cautioned us about seems likely to be at work.