Much discussion has been made regarding Gallup's daily tracking poll and the large lead it projects for Mitt Romney. More and more polling data comes in daily from other organizations that suggests Gallup is an outlier. A deeper review of Gallup's likely voter data and Census voter participation data suggests a flaw in Gallup's likely voter model that specifically underestimates the Hispanic vote (and generally underestimates the non-white vote). With President Obama at +63 with non-white voters, you start to get a sense of how big the error might be.
(continue reading below the fold)
Gallup currently lists likely voter rates for various demographic groups. Demographic groupings like gender, race or national origin, age, and geographic region go into Gallup's weighting of its survey samples of registered voters. These weights attempt to remove the effects of over sampling or under sampling various groups. It is against this weighted group of registered voters that Gallup applies its likely voter model.
Thankfully we have freely available registration and voting data from the Census Bureau that can help us unravel what is going on with Gallup's numbers. The Census Bureau publishes registration and voter participation data by election year. We will compare Gallup's likely voter rate (the likelihood a registered voter will vote) and the Census voter participation rate (the number of self-reported voting participants divided by the number of self-reported registered voters). This Census voter data is further broken out by the same demographic groups Gallup uses in its registered voter weighting.
We begin by assuming Gallup has indeed put together a survey representative of all registered voters, since their registered voter numbers match the overall polling averages. We then evaluate Census data trends over time in an attempt to determine where Gallup's likely voter screen breaks down.
For this exercise, we will compare Gallup's likely voter rates with the Census data from the Presidential election years of 1996 through 2008.
Gallup currently reports an 83% likely voter rate for all registered voters. As a point of comparison, the Census calculated an 89.6% participation rate for registered voters in 2008. This compares to 88.5% in 2004, 85.6% in 2000, and 82.3% in 1996. The Census data shows registered voter participation rates increasing over time. So, we should expect that Gallup's likely voter rate will increase to somewhere around 89-90% by election day.
We are still two week out from the 2012 election, so Gallup's numbers don't currently reflect what the likely voter rates will be on election day. A review of Gallup's 2012 data finds that likely voter rates have been increasing as we near the election. Since April, the white (non-Hispanic) likely voter rate has increased 5 points, the black likely voter rate has has increased 6 points, and the Hispanic likely voter rate has increased a whopping 13 points.
Since Gallup is just a snapshot of a moving target, we can't simply match Gallup's current likely voter rates with Census data of past voter participation rates. To overcome this problem, we will first compare voter participation rates of white (non-Hispanic) voters to all registered voters. Then we will compare voter participation rates of non-white demographic groups relative to the white (non-Hispanic) voter participation rates.
Gallup's likely voter rate for white (non-Hispanic) voters is 86%, 3 points higher than Gallup's likely voter rate for all registered voters. This means Gallup currently reports white (non-Hispanic) voters as 3.6% more likely to vote than all registered voters. The Census calculated an 89.9% voter participation rate for white (non-Hispanic) voters versus an 89.6% rate for all registered voters in 2008, meaning white (non-Hispanic) voters were 0.4% more likely to vote than all registered voters. In the 2004, white (non-Hispanic) voters were 1.1% more likely to vote. In 2000, they were 0.8% more likely to vote, and in 1996 they were, once again, 1.1% more likely to vote.
Gallup's white (non-Hispanic) likely voter rate should continue to increase through election day. But the other demographic groups will see their likely voter rates increase over the next two weeks, too. The data suggests Gallup's white (non-Hispanic) likely voter rate should grow more slowly than non-white likely voter rates in the closing two weeks of the 2012 election.
To compensate for noise created by the changing mix of Gallup's likely voters between now and November 6, we can minimize errors in our estimation of likely voter rates by using white (non-Hispanic) voter numbers as our baseline for comparisons to black voters and Hispanic voters.
This chart represents the difference in black and Hispanic voting rates relative to white (non-Hispanic) voting rates. Any points above the x-axis indicate that the voter demographic had higher voter participation rates than white (non-Hispanic) voters. Any points below the x-axis indicate white (non-Hispanic) voters had higher voter participation rates. Except for 2008, white (non-Hispanic) voters have consistently voted at higher rates than their black and Hispanic counterparts. But you can see how out of line Gallup's data for 2012 is compared to the prior periods.
The likely voter rate for black voters is 85% in the current Gallup survey. Gallup currently shows black voters as 2.3% less likely to vote than white (non-Hispanic) voters. In 2008, black voters were 3.2% more likely to vote than white (non-Hispanic) voters. Black voters were 2.4% less likely to vote in 2004, 2.5% less likely to vote in 2000, and 4.0% less likely to vote in 1996.
Taking all of this into account, we should expect the likely voter rate for black voters in Gallup's survey to fall somewhere between 2.3% less likely to vote and 3.2% more likely to vote than white (non-Hispanic) voters. A 5.5 point swing in relative likelihood to vote is pretty significant.
As you will soon see, Gallup's likely voter rate for Hispanics is even more out of line with recent trends. Gallup reports a 71% likely voter rate for Hispanic voters, meaning Hispanic voters are 17.2% less likely to vote than white (non-Hispanic) voters. In 2008, Hispanic voters were 6.6% less likely to vote than white (non-Hispanic) voters.
Hispanic voters where 8.9% less likely to vote in 2004, 9.0% less likely to vote in 2000, and 9.7% less likely to vote in 1996.
We should expect the likely voter rate for Hispanic voters in Gallup's survey to settle somewhere between 9.0% less likely to vote and 6.6% less likely to vote than white (non-Hispanic) voters. Eight to 11 points represents a huge shift from Gallup's numbers today. Barring a late change in their likely voter methodology, I believe Gallup will continue to underestimate the size of the Hispanic vote up through election day.
In my opinion, the data suggests Gallup's likely voter model is underestimating the non-white proportion of likely voters. With Obama currently standing at +63 among non-white voters, you can see how Gallup has become the 2012 outlier.
As Kos mention earlier today, looking at registered voter data instead of likely voter data should give you a better sense of where the election stands. And hopefully you understand why the ground game for voter registration is so important. The true hurdle to voting is registration. Only 55-60% of all adults vote in a Presidential election. But if you get someone registered to vote, that person in nearly 90% likely to vote. America's apathy toward voting is really an apathy toward registering.
source data:
http://www.census.gov/...
http://www.gallup.com/...
data accessed October 23, 2012
---------------------------------------------------------------------------------------------------
EDITED TO ADD October 24, 2012 1:21PM
I didn't have the exit data I needed at the time of publishing to highlight why the demographic mix of race and national origin is so important in a likely voter model.
source: http://elections.nytimes.com/2008/results/president/exit-polls.html
The chart above plots the demographic mix of race and national origin in Presidential elections since 1980. You can see that the electoral share of white voters has dropped form 89% in 1980 to 74% in 2008, while the electoral share of black voters has increased from 10% to 13% and the share of Hispanic voters going from 2% to 9%. Undercounting a group can throw off a likely voter model, but undercounting a growing group can create a significant error.
I have purposely avoided "unskewing" the Obama-Romney vote split. There is too much room for error because of incomplete data published by Gallup. This exercise was meant to try and explain why Gallup is acting like an outlier this election season.