The news that Nate Silver is leaving the New York Times for a role at ESPN and ABC News (corporate synergies! They're a thing!) has occasioned some interesting posts on what he got right during the election.
The typical answer to this is, well, "the election." But getting the election right was no great feat. The betting markets got the election right. The pollsters got the election right. The polling aggregators, like Real Clear Politics, got the election right. The modelers — which included Silver, but also included Sam Wang and Drew Linzer, among others — got the election right. Wonkblog's election model called the election right — and it did it in June.
The truth is that 2012 just wasn't a very hard election to call. The polling data all pointed in the same direction, even if many pundits refused to believe what it told them. The secret of the modelers — and it's not much of a secret — is that they listened to the polling data. Silver et al got the credit for calling the election right, but the bulk of that credit should really go to the pollsters, without whom none of the modelers could have made any calls at all.
Indeed, one of Silver's incorrect calls came in the North Dakota race, which Wang called correctly. Why did Wang get it right and Silver get it wrong? Because Wang's model stuck even closer to the polls than Silver's model did.
So if all Silver did was build a polling-aggregation model that could call the election correctly on Nov. 1, that wouldn't be much of a trick. Anyone with access to the polls could've done that.
But Silver had two other innovations, both of which are, I think, more important in explaining the appeal — and potential scalability — of his work. The first is that his model begins many, many months before the election, and long before the polls become particularly predictive or frequent. At that point, Silver's model doesn't mainly run on polls (it becomes poll heavy as the election nears and the polls become more predictive). It uses ideology and incumbency and economic growth.
I think of that model as a journalistic innovation more than a statistical one. It gave Silver a way to cover the election at a time when everyone knows the polls aren't worth much but people want to read about the election anyway. It wasn't, however, a very good model. If you look at Silver's November 2011 New York Times magazine story, "Is Obama Toast?", you see a model that's way too pessimistic on Obama's chances.
It concludes, basically, that so long as Mitt Romney is the nominee, "the odds tilt slightly toward Obama joining the list of one-termers." Even in a scenario where GDP growth was an amazing 4 percent in 2012, it gave Obama only a 60-40 shot over Romney. As it was, GDP growth was 2.2 percent in 2012, and yet Obama never fell behind Romney in the battleground states. I think a fair read of the election suggests that Obama's chances were much more robust than Silver's early model indicated.
But if that early model didn't work to predict the election, it served Silver's other, and most important, journalistic strength: narrativizing the data.
The core challenge of covering elections is that pretty much nothing important happens on any given day. That's particularly true 12 months before the election, which is when Silver wrote his magazine story, but it's even true in the days just before an election.
The way a lot of horserace coverage deals with this problem is by blowing up unimportant news — gaffes and ads and the like — into stories that makes the readers feel like they're learning urgent new facts about the campaign even as nothing changed that day and whatever gaffe or ad or speech got made stands almost no chance of influencing the campaign.
There've been election models before Silver's. But their proprietors proudly stood in opposition to this trend. They pointed to their models and said, "See? Most of this stuff doesn't matter, and there's no reason to be covering it." Analytically, they might well have been right about that. But people still wanted to read about the election.
What Silver figured out how to make data-driven election journalism into a daily product that could satisfy political obsessives.
On any given day, you could head to FiveThirtyEight and get a new forecast and an engaging and clear explanation from Silver on what had changed in the forecast. Rather than covering the slow days of the election through the incremental news of the campaign trail, he covered it through the incremental changes on his spreadsheet.
So on Oct. 1, Silver's post was, "New Polls Raise Chance of Electoral College Tie." They didn't raise it that much, of course — it went from 0.3 percent to 0.6 percent, and would fall again in the days to come — but it was a fascinating look at how the electoral data had changed that day. On Sept. 26, he looked at new polls that suggested that Obama might overperform in battleground states.
Every day, the data told a new and interesting story. Sometimes it was a story about something in the news, but more often, it was just a story about the election, and the news peg, so far as there was one, was some bit of movement in Silver's forecast, or in the polls he was watching. And so election obsessives could go to Silver every single day and read something new, even though nothing had really happened.
Silver's reputation as a math wizard often obscures his innovations as a journalist. But it's the latter that makes him such a valuable hire for ESPN and ABC News. Lots of people can run the numbers. But Silver can use those numbers to tell readers an engaging, fast-paced and constantly changing story about subjects they care about. That's a rare talent.