Posted on TheAtlantic.com as Why Demographics Can’t Fully Predict How People Vote
Two-thirds of single women voted for Obama, according to the exit polls taken on the day of the election. On the other hand, the majority—53 percent—of married women went for Romney. With marriage on the decline, figured one pundit,
If [Republicans] are unable to attract the support of more unmarried female voters in future elections, they could face years in the political wilderness.
And Fox News reported:
Marital status was a more significant factor than gender this year. Women, a traditional Democratic voting group, backed Obama by 11 points—about the same as by 13 points in 2008. Even so, married women backed Romney by 7 points (an improvement from McCain’s +3 showing). Men backed Romney (52-45 percent), and married men backed him by an even wider margin (60-38 percent).
But these conclusions are overdrawn, because “unmarried women” are a data category more than a lived social group. Most people who aren’t married will get married. And people who are likely not to get married—such as poor people with fewer marriage prospects—tend to have traditional viewsabout marriage anyway. So what distinguishes unmarried women? Birth control is a common explanation. But almost everyone thinks birth control is okay these days, and although single women might be more worried about access, that’s because they’re less likely to have health insurance. In short, even if some women do have a strong identity as members of the single-women category, most unmarried women are just passing through.
What is grouping good for?
From the large sample interviewed by the major media’s exit polling consortium, we can see that simple demography can make some very strong predictions. At the extreme, with just two data points—race and gender—we can guess how a person voted 96 percent of the time, if she’s a black woman.
But that doesn’t tell us why people voted the way they did. On average, for example, black women also have lower incomes, are younger, and have completed less education than whites—and they have worse healthcare.
Source: National Health Interview Survey.
So did black women vote for Obama because they are proud of him as a black man and identify with his personal experience, or because his policies are more beneficial for them as workers, concerned family members, or medical patients?
And the closer a group gets to the 50/50 middle of the odds breakdown, the harder it is to guess what’s going on from the demographics. When the plane I was on Tuesday night landed, I turned on my iPhone to see who had won Ohio. Waiting for the page to load, I looked around and saw a guy a few rows back, smiling and giving me a thumbs up. He must have noticed I’m a white man, almost two-thirds of whom went for Romney. But did he think I was Jewish, or gay? Maybe it was my rumpled casual-business wear, shaggy hair, and dead-giveaway academic backpack. Or my facial expression.
In real life, even with people we don’t know personally, the cues we get from interactions and behavior explain more than simple demographics, although they all go into the quick mix of judgments we are forced to form on short notice throughout the day. In statistical terms, the basic demographic variables leave a lot of variance unexplained.
These demographic categories are useful for prediction at the group level, as punditry has proved. Show me a room full of randomly-chosen Mormons and I can guess that 78 percent of them voted for Romney. But give me 30 seconds with one of them personally and I might figure out whether she is the one who didn’t.
This paradox is why the “big data” people were so central to the campaigns. If you can get someone to give you a few key bits of information—even just a zip code—and then track them as they wander around the web after leaving your site, your models can be much more powerful than the demographics alone. It’s the equivalent of sizing up someone by their shoes, haircut, and the flight they’re on, rather than just sex and race.