Aug. 5: State Polls, National Polls, and Likely Voters

By NATE SILVER

Good news for polling junkies: pollsters seem to be picking up their pace a bit, and we're finally starting to see some state polls released on the weekends. On Sunday, there were two state polls out, each of which were a bit strange.

One poll was in Indiana, where Rasmussen Reports showed an impressive 16-point lead for Mitt Romney. Indiana has been sparsely polled, in part because the state is one of two in the country (along with North Dakota) that prohibits the use of automated surveys (or “robopolls”). Rasmussen Reports, which usually conducts robopolls, has tried a couple of approaches to work around the problem: Their Sunday poll used live interviewers, while another of their polls of the state in May was conducted online.

The online poll had shown just a 6-point lead for Mr. Romney; it's a bit disconcerting that their different methods are yielding such different results. Nevertheless, there is not much sign that Indiana - which President Obama narrowly won in 2008 - is likely to be competitive this year; the campaigns have expended essentially no resources there. With the new survey, the model gives Mr. Obama just a 6 percent chance of winning Indiana - less than a couple of states, like Missouri and Montana, that he failed to win in 2008.

Another poll out on Sunday, however, showed strong results for Mr. Obama: a SurveyUSA poll gave him a 17-point lead in Washington state, just slightly larger than his margin of victory there in 2008.

Like the Indiana poll, I suspect that this one is a bit of an outlier: there are unlikely to be very many states in which Mr. Obama expands his margin from 2008, and Mr. Obama's polling hasn't been especially strong in neighboring Oregon. Still, the Pacific Northwest was a reasonably competitive region in the elections of 2000 and 2004, and it seems to be trending more and more into the “safe Democr at” column.

The other quirky thing about the SurveyUSA poll is that it was a survey of likely voters rather than registered voters. SurveyUSA's prior polls of Washington state - conducted among registered voters - had shown smaller leads for Mr. Obama, ranging between 9 points and 14.

Typically, the Democratic candidate does worse when polling firms switch to likely voter models, since the demographic groups that are more likely to vote Republican also tend to be more likely to vote, period.

Perhaps we shouldn't make too much of the Washington survey. The gain for Mr. Obama could easily have been caused by statistical variance, rather than the switch to a likely voter model, and Washington state is unusual in that voting takes place by mail, which makes turnout especially high.

Our likely voter adjustment is calibrated solely based on cases where pollsters release both registered voter and likely voter numbers within the same survey - something whic h was not done here. But based on the handful of data that is available, we estimate that Mr. Romney will do about 2 points better in likely voter than registered voter surveys.

This is a moot point, however, if pollsters are already using likely voter models. There is no need to make an inference about what the likely voter numbers will be if the polls are already reporting them directly.

At the state level, most of the recent data - perhaps 80 percent of the polls released - are already using likely voter models. Conversely, about two-thirds of the recent national polls were conducted among registered voters.

I'm going into this detail for a somewhat technical reason. Our model perceives something of a difference between state polls and national polls, with the state polls telling a more favorable story for Mr. Obama.

Part of the reason we perceive this difference is because of the likely voter adjustment. Before the likely voter adjustment, our mod el puts Mr. Obama ahead by about 2 points in the average of national polls (similar to the Real Clear Politics average of a 3- points lead). But most of those national polls were conducted among registered voters. After the adjustment our model applies for this, it takes Mr. Obama's lead in national surveys to be only about half a percentage point instead.

On the other hand, most of the state polls do not require a likely voter adjustment, since they were already conducted on a likely voter basis. And that data, on balance, seems to point to an election in which Mr. Obama has more than a half-point lead. Instead, the state data seems to suggest that he's up 2 or 3 points on Mr. Romney over all - and perhaps by a slightly larger margin than that in the most critical states, which is why the model has started to perceive a bit of an Electoral College edge for him.

So this is why I can't quite agree with Sean Trende's well-researched article that suggests (in contra st to our analysis) there isn't any discrepancy between state and national polls. If you take both sets of numbers at face value, I agree that there isn't too much of a difference. But once you account for the fact that most of the state polls were conducted among likely voters, and most of the national polls among registered voters, there does seem to be a gap.

Some sharp readers may perceive an alternate hypothesis: perhaps Mr. Romney won't do any better among likely voter polls after all? Perhaps we shouldn't be adjusting the registered voter numbers in his favor?

It's possible, I suppose, but it goes against a lot of history suggesting that Republicans almost always do a bit better on likely voter polls in presidential years. In fact, the gap has been quite consistent from year to year, and is typically on the order of 1 or 2 percentage points. Meanwhile, surveys of voter engagement suggest that Republicans are paying a bit more attention to the election so f ar this year, which is usually a good predictor of a likely voter gap.

But it would certainly help if more polling firms reported both registered voter and likely voter numbers within the same survey release - that provides for the most direct type of comparison.

Another selfish plea to pollsters: it would also help if we got more data from noncompetitive states with large populations, since these can affect the popular vote quite a bit even if they don't matter to the electoral math. Large blue states like California and New York have been adequately polled, but large-population red states like Texas, Georgia and Tennessee just haven't been. If Mr. Romney holds a 30-point lead in Texas or something - which is worth about 2 points in the national popular vote given its large population - that would help to explain the apparent state poll and national poll discrepancy.