As a graduate student in economics, I have used Google search data to quantify the cost of racism on President Barack Obama's vote total. I compared the rate at which areas made racist searches on Google to Obama’s vote share, controlling for the vote share of the previous Democratic candidate, John Kerry in 2004. After a large set of robustness checks, I estimated that Obama lost about 4 percentage points of the popular vote (more than 4 million total votes) from racism in both 2008 and 2012. Furthermore, he only gained about 1 percentage point from increased black turnout, and I found little evidence that he gained additional white voters due to his race.
In a recent piece, The New Republic's Nate Cohn took issue with my findings. "Much of the correlation between racist searches and Obama’s performance can be explained by two important trends in political geography that pre-date the first black presidential candidate," he wrote, pointing to "the long term decline in Democratic fortunes in the South and Appalachia" and that "Western states … possess an anti-incumbent streak."
I can see why someone would be skeptical that racism was such a large factor, given that Obama won 53 percent of the electorate in 2008 and 52 percent of the electorate in 2012. But most agree that 2008 was a remarkably favorable year for Democrats, and that Obama's 2012 performance was middling for an incumbent. In addition, I do take seriously the possibility of an "Emerging Democratic Majority" as well as the possibility that Obama is an unusually skilled campaigner who ran unusually effective campaigns. In other words, a candidate with these advantages might have performed significantly better were it not for racism.
As a researcher applying a new data source to a difficult problem, I am always open to alternative explanations for my findings. However, neither of Cohn’s two arguments fits the data. Cohn first hypothesizes that some of Obama's poor performance, relative to Kerry's, in places with the highest racist search rates was the continuation of a long-term trend away from the Democratic Party in those places.
There are four reasons this explanation does not work:
First, if this were true, any Democrat, not just Obama, should have underperformed in such areas. In early 2008, SurveyUSA polled hypothetical match-ups between each of Obama, Hillary Clinton, and John Edwards and various potential Republican nominees. Obama consistently underperformed among whites in areas with higher racist search rates. Clinton did not. Edwards did not.
Second, if areas with higher racist search rates were punishing all Democrats in 2008, relative to 2004, a similar relationship should be seen in House voting patterns. House Democrats should have also underperformed in these areas in 2008, relative to 2004. They did not.
Third, the trend argument makes a clear prediction about what should have happened in 2012. Continuing this hypothesized trend away from Democrats, Obama should have performed worse in 2012 than in 2008 in such areas. He did not.
Fourth, there is a very simple test for this hypothesis. I added controls for long-term trends in Democratic performance. If the trend hypothesis were correct, adding these controls should reduce the correlation that I found. It did not.
Cohn's next hypothesis is that Obama over-performed, relative to Kerry, in the West because the West has a pro-challenger bias. However, Kerry was a challenger in 2004. There was no incumbent on the ballot in 2008. Thus, according to this hypothesis, the West should have preferred Kerry, relative to Obama. This hypothesis would go against finding the correlation that I found.
I claim that Obama lost more than 4 million votes from racism. Cohn says he'll “take the under," calling the number “not even close.” But according to the 2008 and 2010 General Social Surveys, 3 percent of non-black Americans who voted for Kerry in 2004 (about 1.5 million Americans) openly admit to surveys that they will not vote for a black president. Of course, I do not believe that we can trust surveys regarding racism. (This is what makes this question difficult and motivates my use of Google search data.) But, surely, we would suspect that surveys, on balance, substantially underestimate racism. Research consistently finds that people dramatically understate socially unacceptable behaviors to surveys. Fewer than half of non-voters, for example, admit that they do not vote to surveys. Cohn's analysis, however, would seem to require that just about every racist admits this to surveys or even that people, on balance, exaggerate their racism to surveys. This seems unlikely.
Let me be clear about the current finding and its robustness: Racist searches are made with what many would consider shockingly high frequency on Google. Areas with the highest rates of such searches were significantly less likely to support Obama than Kerry. They were also less likely to support Obama than Hillary Clinton and John Edwards. They were not less likely to support 2008 House Democrats compared to 2004 House Democrats. These patterns cannot be explained by gun ownership rates, church attendance, age, education levels, race, trends in Democratic vote shares, various measures of liberalism, or region of the country. After more than two years of studying the topic, I've encountered only one explanation that matches the data: Relatively high racist search rate on Google is indicative of relatively high racism, which leads to substantially fewer votes than an African-American presidential candidate would have otherwise received.
Seth Stephens-Davidowitz is a Ph.D. candidate in economics at Harvard.