1956-2012: Gallup's final pre-election poll got the percentage totals for all candidates correct to within ±5.7% of the total vote in every USA presidential election 1956-2012 (this constitutes 15 elections and 40 predictions in all), Their three worst errors were Clinton in 1992 (prediction=49.0%, actual=43.3%, difference=5.7%), Perot in 1992 (prediction=14.0%, actual=18.9%, difference=4.9%), and Reagan in 1980 (prediction=47.0%, actual=50.8%, difference=3.8%); the remaining 37 predictions all had errors≤3.0%. The median error was 1.4%.
Voter behavior: Some have hypothesized that voters may behave differently when genuinely voting, versus pseudo-voting in polls. E.g. in the real election they might be more strategic. Others have hypothesized the opposite: polls do matter, and voters will strategically lie to pollsters similarly to the ways they strategically lie in their votes. Anyhow, this data shows poll results and real election results are very similar, making the second view seem more correct.
However, there may be evidence that supporters of third party candidates tend to (strategically) vote more against them in the real election, versus what they tell pre-election pollsters. This hypothesis is supported, albeit not very significantly, by this historical evidence:
| 3rd-party Canddt | Official vote | Pollster predictions |
|---|---|---|
| J.B.Anderson 1980 | 6.6% | Gallup=8, CBS/NYtimes=10, NBC/APnews=10, DMI=10 |
| Ron Paul 1988 | 0.47% | ?? |
| Ross H.Perot 1992 | 18.9% | TimesMirrorOct26=19% (but oct8-11 gave 8%), Gallup=14%, NYtimesOct4=7% |
| Ross H.Perot 1996 | 8.4% | NBC/WSJ=9, Pew=8, Reuters=7, ABC=7, Gallup=6, CBS/NYtimes=9 |
| Ralph Nader 2000 | 2.74% | Gallup=4, NBC/WSJ=3, ABC/WashPost=3, Zogby=5, Pew=4, CBS=4 or 5, Harris=5, Battlgnd=4 |
| Ralph Nader 2004 | 0.38% | Real Clear Politics many-poll average=1.0; |
| Gary Johnson 2012 | 0.99% | 3% or 4% (CNN/ORC 7-9 Sept 2012); 1% (Rasmussen); 4.3% (JZ Analytics); 6% (Reason/Rupe); 4% (NBC/WSJ 26-30 Sept 2012); 2% (AP/GfK 30Aug-11Sep) |
| Jill Stein 2012 | 0.36% | 1% or 2% (CNN/ORC 7-9 Sept 2012); 2% (NBC/WSJ 26-30 Sept 2012); 1% (AP/GfK 30Aug-11Sep) |
Pollster errors have both signs for third-party candidates, e.g. the pollsters underestimated Perot's total in 1992, got it right for Perot 1996, and overestimated Nader in 2000 and 2004, Anderson in 1980, and Stein and Johnson in 2012. Note that third party canddates often only get on ballot in some states, a different reason poll percentages to overestimate their votes.
Are pollsters getting better? I see little evidence either that pollsters are now significantly better than Gallup 1956-2012 historical performance, or that there is any improving trend in performance. But the fact that there now are more pollsters allows averaging over many poll results. That might constitute a genuine improvement (it looks that way based on the data I've seen, but not enough such data has appeared for me to feel confident that multi-poll averages work better). E.g. here are the last two elections:
2004: Here are the final pre-election Poll Estimates of the Outcome of the 2004 Presidential Election by 19 polling firms. [Data collected by Michael W. Traugott: The Accuracy of the National Preelection Polls in the 2004 Presidential Election, Public Opinion Quarterly 69,5 (2005) 642-654.] The greatest error was 4.7%, made by Fox News on Bush.
| Firm | Field Period | Bush | Kerry | Nader | Other | Bush Lead |
|---|---|---|---|---|---|---|
| Official Election Result | 50.7 | 48.3 | 0.3 | 0.7 | 2.4 | |
| Harris (Online) | 10/29-11/1 | 47 | 50 | 1 | 2 | –3.0 |
| Battleground: Lake | 10/31-11/1 | 48.6 | 50.7 | –2.1 | ||
| Fox News | 10/30-31 | 46 | 48 | 1 | 5 | –2.0 |
| Marist College | 11/1 | 49 | 50 | 0 | 1 | –1.0 |
| Democracy Corps | 10/29-31 | 47 | 48 | 1 | 3 | –1.0 |
| Gallup | 10/31-11/1 | 49 | 49 | 1 | 1 | 0.0 |
| American Research Group | 10/28-30 | 48 | 48 | 1 | 3 | 0.0 |
| Zogby (as revised in election day afternoon) | 11/2 | 49.4 | 49.1 | 1 | 0.5 | 0.3 |
| Harris (Telephone) | 10/29-11/1 | 49 | 48 | 2 | 1 | 1.0 |
| NBC News/Wall Street Journal | 10/29-31 | 48 | 47 | 1 | 4 | 1.0 |
| ABC News | 10/28-31 | 49 | 48 | 1 | 3 | 1.0 |
| Washington Post | 10/28-31 | 49 | 48 | 1 | 3 | 1.0 |
| CBS News/N.Y. Times | 10/29-11/1 | 49 | 47 | 1 | 3 | 2.0 |
| CNN/USA Today/ Gallup | 10/29-31 | 49 | 47 | 4 | 2.0 | |
| Pew Research Center | 10/27-30 | 51 | 48 | 1 | 3.0 | |
| TechnoMetrica Institute of Policy and Politics (version without allocation of "undecided"s) | 10/30-11/1 | 48.6 | 45.3 | 0.9 | 5.2 | 3.3 |
| Battleground: Tarrance | 10/31-11/1 | 51.2 | 47.8 | 0.5 | 0.5 | 3.4 |
| George Washington University Battleground | 10/31-11/1 | 50 | 46 | 1 | 4 | 4.0 |
| Newsweek | 10/27-29 | 50 | 44 | 1 | 5 | 6.0 |
2008:
|
|
Poll data collected by Dr. Costas Panagopoulos and first published in the essay Polls and Elections: Preelection Poll Accuracy in the 2008 General Elections, Presidential Studies Quarterly 39,4 (Dec. 2009) 896-907. A negative "spread" indicates the poll underestimated Obama's final victory margin; positive means overestimated. The worst prediction error was 4.6% by Rasmussen on McCain. "Real Clear Politics" based on averaging many polls produced a final prediction of Obama=52.1% and McCain=44.5%, off by 0.8% and 1.1% respectively.