2012: The Turnout Model That Was Right
Two campaigns went into 11/06/2012 holding pictures of the electorate that differed by about two points. The gaps between groups were wide enough to make those two points decisive.
Published 08/04/2026
Barack Obama won 332 electoral votes to Mitt Romney’s 206 on 11/06/2012 and took the national popular vote by 3.9 points, 51.1 percent to 47.2 percent. The certified national totals were 65,915,795 to 60,933,504. The two campaigns arrived at election day holding different pictures of who would vote, and the pictures differed by roughly two points of the white share of the electorate. The exit poll settled which one was right.
The standard account of 2012 is a technology story about an election-day application that failed. The evidence points upstream. The decisive error was a measurement choice made months earlier, and the Census Bureau’s own count of who voted shows which side had it right. The field software failure was the same habit in a different department: an untested assumption shipped straight into production.
That position is falsifiable. If the composition of the 2012 electorate had moved the way the losing side expected, with white turnout holding up and minority turnout falling from its 2008 level, the diagnosis would belong to the software. The Census count went the other way.
Three mistakes are visible in 2012 and none of them belongs to one party: a composition that was assumed rather than estimated, a likely-voter screen that selected for enthusiasm and then reported the result as the electorate, and an election-day system whose first real use was the day it mattered.
Modeled electorate composition against assumed composition
The Obama campaign built its 2012 plan on an assumption that white voters would be 72 percent of the electorate, down from 74 percent in 2008, and the national exit poll found 72 percent. The Roper Center publishes the full 2012 exit poll table. Republicans argued through the autumn that stronger conservative enthusiasm would hold the white share at the 2008 level or push it slightly higher. NBC News reported both sides’ assumptions before the election. Pew Research Center put the nonwhite share of the 2012 electorate at 28 percent, up from 26 percent in 2008. Pew published the shift on 11/07/2012.
The white share of the electorate, assumed and measured
What each side expected the white share of the 2012 electorate to be, against the exit poll measurements for 2008 and 2012.
Two points of composition sounds like a rounding error until the group margins are put next to it. In the 2012 national exit poll Romney won white voters by 20 points while Obama won Black voters by 87, Asian voters by 47 and Hispanic voters by 44. Those margins are from the national exit poll. When the groups differ by 60 to 100 points, moving two points of the electorate between them moves the national result by more than the margin in several decisive states. Florida was decided by 0.88 points, Ohio by 2.98 and Virginia by 3.88. State margins come from the certified returns.
Obama’s 2012 margin within each racial group
National exit poll vote margins, with each group’s share of the electorate shown in the label.
The decisive evidence is not the exit poll at all. The Census Bureau’s Current Population Survey, which asks a large national sample whether they voted, found that about 1.7 million more Black voters and 1.4 million more Hispanic voters reported going to the polls in 2012 than in 2008, along with about 550,000 more Asian voters, while the number of non-Hispanic white voters fell by about 2 million. Report P20-568 gives all four figures. The Bureau notes that this was the only instance since 1996 of any race group showing a net decrease in voting from one presidential election to the next. In the same survey, Black citizens voted at a higher rate than non-Hispanic white citizens, 66.2 percent against 64.1 percent, for the first time since the Bureau began publishing rates on that basis in 1996. The Census Bureau announced that crossover on 05/08/2013.
Change in the number of voters by group, 2008 to 2012
Census Bureau Current Population Survey estimates of how many more or fewer people reported voting in 2012 than in 2008.
The two instruments disagree on the exact level and agree on the direction. The Edison exit poll put white voters at 72 percent of the 2012 electorate; the Census survey puts non-Hispanic whites at 73.7 percent of the voting population. The definitions differ, one measures people leaving polling places and the other asks people afterward whether they voted, and self-reported voting is known to run high. Anyone citing a single composition figure for 2012 should say which instrument produced it.
That is the whole argument for treating composition as an output rather than an input. An assumed electorate is a single number a campaign chooses and then cannot see past. A modeled electorate is an estimate with a stated interval that can be revised when new contact data arrives, and that can be reported as a range so that a strategist knows which decisions survive at both ends of it.
What the losing campaign’s internal numbers said
The Romney campaign’s final internal polls, fielded on 11/03 and 11/04/2012 and released after the election, showed Romney ahead by 3.5 points in New Hampshire, ahead by 2.5 in Colorado and tied in Iowa. The New Republic obtained and published the numbers on 11/30/2012. Obama carried New Hampshire by 5.58 points, Colorado by 5.36 and Iowa by 5.81. The certified results are on the record. The campaign’s other released states were closer to the result: its Ohio number had Romney down two against an actual Obama margin of 2.98 points, and its Sunday Pennsylvania number had him down three against an actual 5.38.
Romney’s final internal polls against the certified results
Internal polling fielded 11/03 and 11/04/2012 and released afterward by the campaign, next to the certified margin in each state.
| State | Romney internal poll | Certified result | Size of the miss |
|---|---|---|---|
| New Hampshire | Romney +3.5 | Obama +5.58 | 9.1 points |
| Colorado | Romney +2.5 | Obama +5.36 | 7.9 points |
| Iowa | Tied | Obama +5.81 | 5.8 points |
| Ohio | Romney −2 | Obama +2.98 | 1.0 point |
| Pennsylvania | Romney −3 on Sunday | Obama +5.38 | 2.4 points |
Size of the internal-poll miss, by state
The gap between the Romney campaign’s released internal estimate and the certified margin in each of five states. Assembled from the two series below.
Lining those two series up shows where the error concentrated. The internal estimate was 9.1 points away from the certified margin in New Hampshire and 7.9 points away in Colorado, against 2.4 in Pennsylvania and 1.0 in Ohio. The two largest misses were in the two states the campaign was publicly describing as moving toward it in the final days.
Neil Newhouse, the campaign’s pollster, told The New Republic that the biggest flaw was a failure to predict the composition of the electorate, which turned out younger, less white and more Democratic than the campaign expected, and that the Colorado Latino vote had been extraordinarily difficult to poll. The reporter, Noam Scheiber, pushed back on how much the demographic account could carry, noting that a tied Iowa poll would have required Obama to win nearly every undecided voter in the state. Both readings are in the same piece and both are worth holding.
A mechanism is visible in the published detail. Those internal polls asked respondents to rate their interest in the election from 1 to 10, and people who rated themselves 8 to 10 made up roughly 80 to 90 percent of the samples, reaching 88 percent in New Hampshire. Romney ran notably better among the most interested voters. A screen that heavy is a turnout model wearing the clothes of a filter: it answers the question “what if the electorate looks like the most engaged slice of this sample” and then the answer gets written down as the horse race.
The failure mode is quiet. A strict screen produces a conditional estimate, the condition is the composition of the electorate, and the composition of the electorate was the exact thing in dispute. Reporting one number under one screen hides the dispute inside the result. Running the same respondent set through two or three defensible screens and publishing all of them internally would have put the disagreement on the table in October, where somebody could still act on it.
The winning side treated the same quantity as something to be re-estimated. The Obama campaign ran the election 66,000 times every night to produce a win probability for each state and allocated the next day’s resources from that readout, and in the final month it held data on about 29,000 people in Ohio, nearly half of one percent of the state’s voters. Michael Scherer reported both figures in Time on 11/07/2012. Sample size is the smaller part of that difference. The larger part is that one side’s electorate was an output that moved every night and the other side’s was an input fixed early.
How much Ohio data the winning campaign was holding
Respondents behind the Obama campaign’s Ohio estimates in the final month, against the size of a routine public state poll.
The field technology failure on election day
Project ORCA, the Romney campaign’s election-day voter-tracking system, went live at 6 a.m. on 11/06/2012 and then went down for roughly 90 minutes, which the campaign attributed to an overload of its data servers in Boston. NPR reported the outage and its causes on 11/12/2012. It was presented to volunteers as an app and was in fact a mobile-optimized website, so volunteers who went looking for it in the Apple and Android stores found nothing. Roughly 800 volunteers at headquarters were allocating resources in real time, and the organizer John Ekdahl described tens of thousands of volunteers in the field unable to log in or add voters, and said his questions about stress testing, redundancy and denial-of-service protection had been brushed aside.
The campaign’s digital director, Zac Moffatt, said afterward that “without a doubt, ORCA had its challenges,” while maintaining that the system worked and recording that it captured data on 14.3 million voters and 5,397 reported ballot box incidents.
No serious account claims the outage decided the presidency. Obama’s margins in the states that mattered were far larger than any plausible field effect in a single morning. The portable lesson is about release discipline. A system whose first production use is the one day it is needed has never been rehearsed, and rehearsal is the only thing that finds a load problem. A full-scale dry run with the actual volunteers on the actual hardware costs a weekend in October and is the cheapest insurance a campaign can buy.
What a campaign can take from 2012
Three practices follow directly, and none of them belong to one party. Treat the composition of the electorate as a modeled quantity with a stated interval, written down by group and by geography, and revisit it as contact data accumulates. Report horse-race results under at least two likely-voter definitions, because a single screen is an assumption that disappears into the number it produces.
And rehearse the election-day system at full scale before election day, including the people who will use it. The first two practices are about knowing what you believe. The third is about finding out whether the thing you built can carry the weight you are about to put on it.
Frequently asked questions
What did the 2012 exit poll say about the composition of the electorate?
The 2012 national exit poll found that white voters were 72 percent of the electorate, Black voters 13 percent, Hispanic voters 10 percent and Asian voters 3 percent. Pew Research Center put the combined nonwhite share at 28 percent, up from 26 percent in 2008. Barack Obama won Black voters by 87 points, Asian voters by 47 and Hispanic voters by 44, while Mitt Romney won white voters by 20.
What did Mitt Romney's internal polls show in 2012?
The Romney campaign's final internal polls, fielded on 11/03 and 11/04/2012, had Romney ahead by 3.5 points in New Hampshire, ahead by 2.5 points in Colorado and tied in Iowa. Barack Obama won all three states by between 5.36 and 5.81 points. The campaign's pollster Neil Newhouse attributed the error mainly to a failure to predict the composition of the electorate.
What was Project ORCA in the 2012 election?
Project ORCA was the Romney campaign's election-day system for tracking which supporters had voted. It went live at 6 a.m. on 11/06/2012, was a mobile-optimized website rather than the app volunteers were told to expect, and went down for about 90 minutes when the campaign's Boston data servers were overloaded. The campaign's digital director said it recorded data on 14.3 million voters.
Why was the Obama campaign's 2012 turnout model more accurate?
The Obama campaign treated the composition of the electorate as an estimate it revised continuously rather than as a fixed assumption. It ran 66,000 simulations of the election each night to produce state win probabilities, allocated the next day's resources from that output, and in the final month held data on about 29,000 people in Ohio alone.
More from the blog
Every campaign has a number for who will vote, and most campaigns have never written it down where anyone can argue with it. Our Modeling work produces that composition as an estimate with an interval, by group and by precinct, and reruns it as contact data comes in. Our Bespoke Tech Tools work builds the election-day systems that carry it, and rehearses them at full scale before the day. Tell us which race you are working and we will come back with a written scope.