The Blue Wall that wasn't: state polls and the 2016 miss
Abstract. National polls had Hillary Clinton about 4 points ahead and she won the popular vote by 2.1, a normal-sized miss, but state polls overstated her by 5 to 7 points in Wisconsin, Michigan and Pennsylvania, all of which Donald Trump won. AAPOR blamed late deciders breaking for Trump and state polls that did not weight by education.
Summary
The final national polling averages had Hillary Clinton ahead of Donald Trump by 4.0 points on the mean of six published aggregators. She won the popular vote by 2.1 points (48.2% to 46.1%), so the national miss on the lead was 1.9 points, a normal-sized error by historical standards. The damage was done in the states: polls showed Clinton narrowly ahead in Wisconsin, Michigan and Pennsylvania, and Trump carried all three and the Electoral College. AAPOR's 2017 evaluation blamed a late swing to Trump, state polls that did not weight for education, and a smaller turnout shift toward rural and Republican counties.
What the polls said
National polls agreed that Clinton was ahead, by somewhere between 1 and 6 points. The table shows the final polls of the larger national pollsters as stored in our database (four-candidate question where one was asked, so Gary Johnson and Jill Stein take a share).
| Pollster | Fieldwork end | Clinton | Trump | Lead |
|---|---|---|---|---|
| YouGov/Economist | 7 Nov | 45 | 41 | Clinton +4 |
| ABC News/Washington Post | 6 Nov | 47 | 43 | Clinton +4 |
| Fox News | 6 Nov | 48 | 44 | Clinton +4 |
| Monmouth | 6 Nov | 50 | 44 | Clinton +6 |
| NBC News/SurveyMonkey | 6 Nov | 47 | 41 | Clinton +6 |
| Rasmussen | 6 Nov | 45 | 43 | Clinton +2 |
| Bloomberg/Selzer | 6 Nov | 44 | 41 | Clinton +3 |
| Morning Consult | 5 Nov | 45 | 42 | Clinton +3 |
| McClatchy/Marist | 3 Nov | 44 | 43 | Clinton +1 |
The two-way averages published on election eve were Clinton +3.2 at RealClearPolitics (46.8 to 43.6), +3.1 at the New York Times, +3.6 at 270toWin, +4.0 at the BBC, +4.9 at TPM and +5.3 at HuffPost Pollster. Their mean was Clinton 47.3, Trump 43.3, a lead of 4.0 points.
The state averages told a closer story. Eight states holding more than a third of the electoral votes needed to win had polls within 3 points, and AAPOR quotes RealClearPolitics' state averages giving Clinton a 272–266 Electoral College lead. The final polls in the three "Blue Wall" states all had her ahead: on our database's final-poll averages, by 6.4 points in Wisconsin, 6.3 in Michigan and 4.3 in Pennsylvania.
What happened
According to the Federal Election Commission's official totals, Clinton won 65,853,514 votes (48.18%) and Trump 62,984,828 (46.09%). Trump won 304 electoral votes to Clinton's 227, after seven faithless electors.
Against the aggregator average, Clinton's share was understated by 0.9 points and Trump's by 2.8. The national lead was off by 1.9 points. Almost all of the error came from Trump's side: undecided and third-party voters in the polls broke toward him.
The states were a different matter. Trump won Wisconsin 47.2% to 46.5%, Michigan 47.5% to 47.3% and Pennsylvania 48.6% to 47.9%. Together those 46 electoral votes decided the election, and the polls had the wrong winner in each.
Why it happened: what the pollsters found
The American Association for Public Opinion Research set up an ad hoc committee in spring 2016, chaired by Courtney Kennedy of Pew Research Center, before anyone knew the result. Its report, An Evaluation of 2016 Election Polls in the U.S., was released on 4 May 2017. Its findings:
- National polls were fine. Final national polls had an average absolute error of 2.2 points on the margin and a signed error of 1.3 points toward Clinton, among the most accurate since 1936 and well under the long-run average of 4.4 points.
- State polls had a historically bad year. Across 423 state polls in the final 13 days, the average absolute error was 5.1 points and the signed error 3.0 points toward Clinton. In the four previous elections the absolute error had ranged from 3.2 to 4.6. Wisconsin had the largest signed error (6.5 points); Ohio was 5.2; Minnesota, Pennsylvania and North Carolina 4 to 5; Michigan and New Hampshire 3.5.
- Late deciders broke for Trump. In the exit polls, 13% to 15% of voters in Michigan, Wisconsin, Pennsylvania and Florida said they decided in the final week. They went for Trump by nearly 30 points in Wisconsin, 17 in Pennsylvania and Florida and 11 in Michigan.
- Many state polls did not weight for education. College graduates are more likely to answer surveys, and in 2016 education was strongly tied to the vote. Polls that did not correct for it overstated Clinton. The University of New Hampshire's final poll had Clinton up 11; reweighted by education, it showed a tie, close to her 0.4-point win.
- Turnout shifted toward Republican counties, which polls modeled on 2012 turnout missed, though the committee found only small effects from likely-voter models.
- No evidence for "shy Trump" voters beyond some respondents who revealed a Trump vote only after the election, which could equally be late deciding.
The committee also noted that the fall of 2016 had a heavy run of late news, including FBI Director James Comey's letter on 28 October.
What our data shows
Our database holds 20 national pollsters' final polls from the last two weeks. Their average put Clinton 3.5 points ahead, 1.4 points above the certified margin, and 14 of the 20 overstated her lead. That is a modest national miss, of the same size as 2012's error in the other direction (polls then understated Barack Obama by 2.9 points on our numbers).
The state picture is where 2016 stands out. Averaging each pollster's final poll in each state over the last three weeks, polls in the 36 states we hold overstated Clinton's margin by 3.3 points on average, with an average absolute error of 5.2 points. In 2008 the signed error was close to zero (-0.3) and in 2012 it ran 1.5 points toward Mitt Romney. The errors also followed a clear pattern: they were largest in white, working-class Midwestern states.
| State | Final-poll average | Result | Error on lead |
|---|---|---|---|
| Wisconsin | Clinton +6.4 | Trump +0.8 | 7.2 |
| Iowa | Trump +2.4 | Trump +9.4 | 7.0 |
| Michigan | Clinton +6.3 | Trump +0.2 | 6.6 |
| Ohio | Trump +2.6 | Trump +8.1 | 5.5 |
| Pennsylvania | Clinton +4.3 | Trump +0.7 | 5.0 |
| North Carolina | Clinton +0.4 | Trump +3.7 | 4.1 |
| Arizona | Trump +3.4 | Trump +3.5 | 0.1 |
| Nevada | Clinton +0.5 | Clinton +2.4 | -2.0 |
Arizona, Georgia and Texas were close to right, and in Nevada, with its large Hispanic electorate, the polls understated Clinton. A single national error cannot produce that spread. It fits AAPOR's account: an education gap that unweighted samples missed, concentrated where non-college white voters were a large share of the electorate.
What the pollsters changed
The main fix was weighting by education. When the 2017 report came out, Pew Research Center summed up the state polls of 2016 this way: "more often than not, the polls were not being adjusted for education." Four years later, AAPOR's 2020 task force found that 92% of the state-level presidential polls in the final two weeks that disclosed their adjustments accounted for education. By 2024, 92% of the firms AAPOR surveyed said they always weighted on education.
Other changes followed. SurveyMonkey's Mark Blumenthal, a member of the committee, reported that separating postgraduates from other college graduates in the weights would have cut Clinton's lead in its final tracking poll by half a point. The committee asked for more high-quality state polling, and the volume shifted: by 2024 there were nearly twice as many state-level presidential polls as national ones, most of them in the seven swing states. Pew found that 26% of pollsters changed how they sampled or interviewed people between 2016 and 2018.
What it means for our forecast
2016 is the first of three presidential elections in a row in which polls overstated the Democrat, and it is part of the history our US model learns from. We correct polls by half the average past miss, and keep that correction only if it would have improved past forecasts made for each election without knowing its result. That means a 2016-style error is partly built into the central estimate rather than treated as a surprise.
The bigger lesson of 2016 is about correlation. The states missed together, in the same direction, with the biggest errors in the Upper Midwest. Our simulation replays past state-level errors as whole patterns, so when one Rust Belt state misses, its neighbors tend to miss with it. A forecast that treated Wisconsin, Michigan and Pennsylvania as independent coin flips would have made Clinton look far safer than she was. The method is set out in our US simulation paper, and the record of individual firms is in our US pollster ratings.
Sources
- AAPOR: An Evaluation of 2016 Election Polls in the U.S. (Ad Hoc Committee on 2016 Election Polling, 2017)
- AAPOR: Task Force on 2020 Pre-Election Polling (2021)
- Federal Election Commission: Federal Elections 2016
- American Presidency Project: 2016 General Election Preference for President (RealClearPolitics Average)
- Pew Research Center: Q&A: Political polls and the 2016 election (4 May 2017)
- Pew Research Center: How Public Polling Has Changed in the 21st Century (19 April 2023)
- Wikipedia: Nationwide opinion polling for the 2016 United States presidential election (aggregator table, cross-check)
- COSSA: AAPOR Report Assesses 2016 Election Polling Performance
How to cite
Polling Forecast (2026). The Blue Wall that wasn't: state polls and the 2016 miss. Retrieved 2 November 2026 from https://www.pollingforecast.com/research/missed-us-2016. Live figures change as new polls arrive; the date above is when the text was last revised.