The Science of Election Estimation

I've been having a Twitter exchange with James Eiden about how you can call an election with such small percentages of the votes in for a state.  I didn't think it's that difficult to figure out how it's done, but apparently, it needs more than 140 characters so here goes:

News organizations triangulate from a number of different sources: primary voting, new voter registrations, previous elections, demographics, and yes, exit polling.  Exit polling, given its unpredictability, I imagine is mostly used as a backup check at this point--to make sure long tail events didn't occur. 

So how do you call a state with 1% in?

Let's take our example state of Jesusland:

Jesusland has 100 voting districts, each representing an equal number of voters.


What are things we know about Jesusland before the election happens?

- Number of registered voters for each party in each district. 

- Primary results from that year.

- Historical records.


We know most people generally vote along party lines, so when you have districts that are heavily Republican or Democrat, you can pretty much count on those folks to go to one side or the other right away.  This is where exit polling comes in.  When you exit poll all of Jesusland's Republican heavy election districts and it seems like there wasn't a major shift against the tide, you can count those votes as in the bag within a margin of error--especially if those same districts turned out in strong numbers for these same candidates in Jesusland's primaries.

That may leave only about a third of the population of Jesusland really in play, making each district about 3.33% of the contested group.  To put the statistical significance of that number in perspective, there are about 180 million potential voters in this whole country.  When national polls are conducted, they're done with less than 10,000 voters and statisticians put them at a 3.5% margin of error.  To get 3.33% of votors responding would be like getting 6 million people nationally to answer a Gallup poll.  So when you see 1% of people reporting and you think it's too early to call, keep in mind that's orders of magnitude more than Gallup usually gets and they're able to get within a few percantage points.  When you get that large of a percent in key districts that are highly contested, you can pretty much call the rest of the race within fair degrees of certainty assuming no major cross-party upheavals.

I mean, it's no different than the whole map of the US, really.  If Ohio, Florida, and Pennsylvania polls closed at 4 in the afternoon, we'd know who the president was by dinnertime.