Voters in a dozen states are weighing in on who they want to be their parties' nominee, and if the polls are right, it seems likely to be a very good night for Republican Donald Trump and Democrat Hillary Clinton. But are the polls right?
I've written extensively about how little you can really learn from primary polling and how difficult polling in general has become in the modern world. There are good reasons to be skeptical that we're getting an accurate read on the true electorate in any given attempt to measure voter attitudes. Yet one of the best ways to overcome problems like random error and "house effects" (that is, the ways a polling firm's biases and methodological quirks influence its numbers) is to look at what's happening across a number of different surveys taken by a number of different outfits.
Right now, the consensus is that Trump and Clinton look strong. An online survey of 1,254 Republicans conducted for Bloomberg Politics recently found Trump leading his rivals with 37 percent of the vote in the Super Tuesday states; Florida Sen. Marco Rubio and Texas Sen. Ted Cruz received 20 percent apiece. A smaller national telephone survey from CNN/ORC released yesterday also had Trump ahead by double digits. Meanwhile, the blog FiveThirtyEight views Clinton as the overwhelming favorite in Alabama, Arkansas, Georgia, Massachusetts, Tennessee, Texas, and Virginia.
If any poll average is going to correctly predict the outcome in a particular race, it will be the poll average as it stands right before voters actually cast their ballots. When the National Council on Public Polls calculates how far off the industry was in a particular year, it uses the final poll from each outlet to do it—because polling becomes more predictive the closer you get to Election Day. As I've observed previously, RealClearPolitics' David Byler has crunched the numbers and found that, in fact, during the last three cycles, polls' "predictive power" increased from about 0.6 out of 1.0 shortly after Thanksgiving to well above 0.9 out of 1.0 at the time of the first primaries and caucuses.
But there are real problems with the current state-specific polling averages. Namely, they're averaging an awfully small number of data points. In Virginia, RCP is making use of a grand total of four surveys taken since January 1 on the Republican side. In Alabama, there have been just three. The only poll of Arkansas GOP voters conducted in 2016 happened nearly a month ago.
Texas and Georgia are a little bit better. But consider that in Iowa, more than two dozen surveys were conducted in just the month leading up to the GOP caucuses, including a couple that finished fielding on January 31, one day before voting occurred—and pollsters still missed the scale of Cruz's last-minute surge. (The final GOP RCP average in the state had Trump up by nearly 5 percentage points. Instead, he lost by 3.)
There are admittedly some places where one candidate is so far ahead it seems almost inconceivable that anyone else could win. But in states where minimal polling has taken place, even a big lead is not a sure bet—a lesson we should have learned well from Eric Cantor. The former Virginia congressman was, you may recall, expected to easily win re-election in 2014. An internal poll his campaign took in late May of that year reportedly found him crushing his primary opponent by better than a 2–1 margin; two weeks later, he was defeated by a stunning 11 points.
Primary polling is harder than general election polling for a variety of reasons. For one thing, voters are more likely to change their mind at the last second. Most people are pretty firmly aligned with one party or the other, and in November they're highly likely to vote the way they did in previous years. But in a nominating contest, all the candidates are (ostensibly) members of the same party, so a voter has to look deeper in trying to decide whom he or she likes best. This makes their preferences more fluid.
For another thing, turnout is a lot lower in off-year and primary elections than in presidential generals, and that makes separating the wheat (people who will actually turn out on Election Day and thus people whose opinions you want in your survey) from the chaff (people who say they plan to vote but actually won't, and who should ideally be excluded from your survey) simultaneously more difficult and more important to getting the call right.
And then there's the fact that state-wide surveys often involve fewer interviews than national ones. Consequently, they have larger margins of error.
Of course, just because a poll has a high margin of error doesn't mean it's necessarily wrong. Just because a poll was taken days or weeks before the voting happens doesn't mean people necessarily decided to support someone else in the interim. And when a lot of different polls from a lot of different pollsters are all pointing in the same direction—in this case, toward a good night for Trump and Clinton—the odds are that they're right.