Method and data
Two surveys, one equation, then the official count.
Amelie and Eric
The American National Election Studies (ANES) is a probability sample, meaning people were chosen at random with a known chance of selection. They were interviewed at home before the election and again after. It is not the exit poll. The Cooperative Congressional Election Study (CCES) is YouGov’s volunteer panel. About 50,800 people finished the pre-election interview. YouGov ran one search per target person, described below, and kept 32,800.
1. Draw target rows
A target row is one real person-record from the 2006 American Community Survey, a large Census Bureau survey. Each target sits inside an age × race × gender × education cell, meaning one combination of those traits. Registration and past turnout are copied in from the Current Population Survey. Religion, party, and news interest are copied in from the Pew Religious Landscape Study. The frame, the population the targets stand for, is U.S. citizens. It includes people who do not vote.
2. Keep the closest interview
YouGov gives every finished interview a distance score against that one row. A higher score means a worse match. It keeps the interview with the lowest score. A race mismatch costs 10. A gender mismatch costs 1.5. Thirty years of age difference costs 1. A party mismatch costs 1.5. A registration mismatch costs up to 4. The closest interview is kept even when the match is weak.
3. Weight, then rescale
A weight sets how much each interview counts. First, a logistic model, a regression for a yes-or-no outcome, uses age, education, gender, and turnout. Next, gender, race, education, and age are forced back to the citizen frame. Any weight above 7 is cut. Last, all weights are divided by the same constant so the average weight is 1. This does not change any percentage or slope.
The shared result
The shared model uses only the traits both files contain: party, age, and race. The party slope is how much the vote score changes for each step toward Republican on the party scale. In the shared model, the party slope is the same number in both surveys.
Figure 1
Party coefficient in the shared model
N = 23,681 · SE 0.002
N = 1,538 · SE 0.010
N = 25,814
N = 1,436
Each bar is the party coefficient: how much the vote score changes for one step toward Republican on the party scale. The vote score is +1 Obama, −1 McCain, 0 another candidate. The shared model uses only party, age, and race, the traits both surveys contain. Bar length is the size of the negative slope. The scale runs from 0 to −0.40. N is the number of respondents. SE is the standard error, a measure of how precise the estimate is. Source: rerun bars from this replication; published bars from Table 5 of Ansolabehere and Rivers (2013).
The other shared slopes are close. A covariate, in the table below, is a trait the model holds constant. Hispanic is the one gap, and the article already flags it. Turning the CCES weights off moves the party slope only from −0.317 to −0.320. So the relationship is already in the interviews.
| Covariate | CCES | ANES |
|---|---|---|
| Party, 1 to 7 | −0.317 | −0.317 |
| Age, per 10 years | −0.040 | −0.050 |
| Black | +0.260 | +0.325 |
| Hispanic | +0.063 | +0.159 |
| Other race | +0.014 | +0.259 |
The two-party share counts only people who chose Obama or McCain. Among them, the weighted Obama share is 53.9% in the CCES and 54.9% in the ANES. The state check compares the CCES with the official vote in each of the 50 states. The CCES is +1.0 point too Democratic on average. The typical state error is 3.4 points.
Figure 2
Obama share of the two-party vote
Each bar is the weighted percent who chose Obama, out of everyone who chose Obama or McCain. Votes for other candidates are left out. The bar scale is 0 to 100 percent. Source: this replication, using the weighted 2008 CCES and ANES. For comparison, the article’s exit poll, a survey of voters leaving the polls, puts Hispanic voters at about 67 percent Obama.
Figure 3
CCES state vote minus the official two-party share
| Mean error, this rerun | +1.00 points |
|---|---|
| RMSE, this rerun | 3.43 points |
| Mean error, article | +0.27 points |
| RMSE, article | 2.43 points |
Each state error is the CCES two-party Democratic share minus the certified state count, in percentage points. Positive means the survey is more Democratic than the official result. Mean error is the average across fifty states. RMSE, the root mean square error, is the typical size of a state error, ignoring direction. Source: rerun rows from this replication; article rows from Ansolabehere and Rivers (2013).
Both surveys record what people said about voting. Neither checks the voter file, the official record of who voted. After weighting, 70.7% of the CCES say they voted. Actual turnout among eligible citizens in 2008 was about 62%.
Sources
- Ansolabehere and Rivers, “Cooperative Survey Research,” Annual Review of Political Science (2013)
- 2008 Cooperative Congressional Election Study (CCES), Harvard Dataverse
- 2008 ANES Time Series, American National Election Studies
- American Community Survey, U.S. Census Bureau
- Voting and registration, Current Population Survey, U.S. Census Bureau
- Religious Landscape Study, Pew Research Center
- YouGov, the survey firm that runs the CCES panel
- Federal election results, Federal Election Commission
- 2008 general election turnout, United States Elections Project