The particulars of polls
As a recovering pollster (I worked for Democratic pollster Peter Hart from 1974 to 1981), let me weigh in on the controversy over whether the polls are accurate. Many conservatives are claiming that multiple polls have overly Democratic samples, and some charge that media pollsters are trying to discourage Republican voters.
First, some points about the limits of polls. Random sample polling is an imprecise instrument. There’s an error margin of 3 or 4 percent, and polling theory tells us that one out of 20 polls is wrong, with results outside the margin of error. Sometimes it’s easy to spot such an outlier; sometimes not.
In addition, it’s getting much harder for pollsters to get people to respond to interviews. The Pew Research Center reports that it’s getting only 9 percent of the people it contacts to respond to its questions. That’s compared to 36 percent in 1997.
Interestingly, response rates are much higher in new democracies. Americans, particularly in target states, may be getting poll fatigue. When a phone rings in New Hampshire, it might well be a pollster calling.
Are those 9 percent representative of the larger population? As that percentage declines, it seems increasingly possible that the sample is unrepresentative of the much larger voting public. One thing a poll can’t tell us is the opinion of people who refuse to be polled.
Then there is the problem of cellphone-only households. In the 1930s and 1940s, pollsters conducted interviews in person because half of households had either no phone or (your grandparents can explain this) a party-line phone.
By the 1970s, phone ownership was well nigh universal, and pollsters mostly phased out in-person interviewing. Phone interviews are much cheaper and quicker.
But today the percentage of households without landline phones is increasing. Under federal law, cellphone numbers have to be hand-dialed rather than dialed by computer, as landline numbers are now even when live interviewers ask the questions.
Cellphone-only individuals tend to be younger and more Democratic than landline owners. Most pollsters are conducting a set number of interviews with cellphone-only households. But they can only guess at what percentage of the electorate they’ll constitute. Oversample them, and you’ll get overly Democratic results.
That, many conservatives are arguing, is what pollsters have been getting in polls this month. They point out that Mitt Romney is running ahead among Independents in many polls but trails overall.
This can only happen if Democrats have a big lead in party identification, as they did in 2008. In the exit poll then, 39 percent of voters identified themselves as Democrats and 32 percent as Republicans.
In contrast, exit polls showed an even break on party identification in 2004 and 2010. But many September polls and some earlier polls showed Democrats with an even bigger party identification lead than four years before.
That seems implausible. Party identification does change over time, as exit polls indicate. But it usually shifts gradually rather than suddenly, as current polls suggest.
There is evidence that since the Charlotte convention Democrats have become more motivated to vote and have narrowed the advantage in enthusiasm Republicans have had since 2010. In that case, more Democrats may be passing through screening questions and getting polled.
I don’t believe that any of the media pollsters have been tilting their results in order to demoralize Republicans, though I do look with suspicion on the work of some partisan pollsters.
But I do have my doubts about whether samples with more Democratic party identification than in 2008 are accurate representations of the actual electorate. Many states with party registration have shown big drops in registered Democrats since then.
Pollster Scott Rasmussen, who weights his robocall results by party identification, adjusted monthly, has shown a much closer race than most pollsters who leave party identification numbers unweighted. So has the Susquehanna poll in Pennsylvania.
It may be that we’re seeing the phenomenon we’ve seen for years in exit polls, which have consistently skewed Democratic (and toward Barack Obama in the 2008 primaries). Part of that is interviewer error: Exit poll pioneer Warren Mitofsky found that the biggest discrepancies between exit polls and actual results were in precincts where the interviewers were female graduate students.
But he also found that Democrats were simply more willing to fill out the exit poll. That raises the question: Are we seeing the same thing in this month’s polls?