How Much Did Racial Prejudice Affect the 2008 Election?

I’ve done a simple analysis that speaks to this question. Using the 1972-2008 American National Election Studies, I examined only white voters and modeled presidential vote (excluding third parties) as a function of three factors:

  • Identification with a party, which is coded in 7 categories ranging from strong Democrat to strong Republican. This is a conventional measure in political science.
  • Self-reported ideology on the liberal-conservative spectrum, which is also coded in 7 categories ranging from strong liberal to strong conservative. This is also a conventional measure, although hardly the definitive measure.
  • A measure of affect toward blacks. Beginning with the 1972 survey, respondents were asked to evaluate various groups on a 1-100 “feeling thermometer,” where 1 signifies very cool feelings and 100 very warm feelings. I subtracted affect toward blacks from affect toward whites. Thus, people with higher scores evaluated whites more favorably than blacks. (Computing the difference between evaluations of whites and blacks helps deal with scale incomparability across respondents; see Lee’s felicitously titled article.)

(Nerdy details: Each factor is scaled 0-1. Presidential vote is coded 1-Republican and 0-Democrat. The model is a logit model, with survey weights. This is particularly important in 2008, when the ANES intentionally sampled a larger number of blacks.)

Yes, this is a simple model. Yes, I could control for other things. Yes, it’s not a perfect measure of prejudice. (And so forgive me the title of the post. I’m just trying to draw some attention amidst the other interesting stuff on people’s RSS feeds.) To summarize: this is a blog post, not a journal submission.

Below I plot the coefficients and confidence intervals from a model in each presidential election year. Larger values mean a stronger relationship between the factor in question and presidential vote. If the confidence interval does not cross 0, then the coefficient attains conventional levels of statistical significance.


The results show the increasing effects of both party identification and ideology. These are familiar findings (see, e.g., here).

Affect toward blacks had a statistically significant impact in 1972-92, but its impact was much closer to 0 in 1996 and 2000. Its effect was larger in 2004 than in the two previous elections, although not statistically significant. In 2008, however, its effect was larger still and statistically significant.

What does the 2008 effect, substantively speaking? It means that, controlling for party identification and ideology, the more negatively a white person evaluates blacks, relative to whites, the more likely they are to vote for McCain. For example, a white person who calls him or herself an independent and a moderate and evaluates whites and blacks equally would have a .51 probability of voting for McCain. A person who is an independent and a moderate, but who evaluates blacks 20 points lower than whites on the feeling thermometers, would have a .57 probability of voting for McCain—a 6 point increase.

That is a hypothetical effect at the level of an individual voter. What about at the level of the electorate as a whole? To simulate this effect, I calculated Obama’s expected share of the vote based on predictions from the model above. As before, I am focusing on white voters. Then I simulated his share of the white vote if everyone who had a less favorable view of blacks, relative to whites, were magically transformed into someone with equivalent views of blacks and whites. So what would happen to Obama’s vote share among white voters in this scenario, relative to reality?

It would have increased by 1 percentage point. That’s it.

Sound small? This is because those whose affect toward blacks changed in my simulation may have been predisposed to vote for McCain because of party or ideology. In sum, there are relatively few voters who are both negatively disposed toward blacks and not already strongly committted to a candidate for other reasons.

This analysis buttresses Charles Franklin’s analysis of a pre-election poll, which I linked to in late October. He also found small aggregate effects. Neither of our analysis speaks at all to Electoral College votes, which would require separate analyses of survey data in individual battleground states.

Let me be clear: this analysis is not the final word. Not at all. By any stretch. At all. Seriously. But I am willing to bet that truly sophisticated analyses will show something similar: a significant individual-level effect of racial prejudice on presidential voting in 2008, but not necessarily large aggregate effects on the actual outcome of the election.

7 Responses to How Much Did Racial Prejudice Affect the 2008 Election?

  1. Referee #1 March 10, 2009 at 4:09 pm #

    While the author makes some interesting points about the effect of race, ideology, and party, the graphs would be more persuasive if first differences (of, say, min vs. max) rather than raw logit coefficients were plotted. This would shed light on the relative substantive impacts of the variables, both over time and across variables.

  2. Anonymous March 10, 2009 at 4:16 pm #

    Of course, the question becomes how much does racial bias (which there’s every reason to suspect a feeling thermometer of measuring poorly) possibly affect partisanship or ideology? Or the reverse? What are the indirect effects?

    I’m having trouble thinking of a way of teasing out both the direct and indirect effects, because I can’t think of any instruments, and a block recursive approach imposes a time order on this that I’m not happy with (ie, ANY time order).

  3. John Sides March 10, 2009 at 4:37 pm #

    Referee #1: Please. I am a lazy, pajama-clad, parents’ basement-dwelling blogger.

    Anonymous: Perhaps I’m underestimating the effect of racial prejudice by not accounting for indirect effects. But, like you, I’m unsure of how to estimate those effects. And I’m somewhat in agreement on the value of feeling thermometers. Presumably a better measure of prejudice would have stronger individual-level effects. But again, I wonder if the aggregate effects would be much greater than those I report here.

  4. Andrew March 10, 2009 at 8:39 pm #


    1. I like your use of the secret weapon plot.

    2. I don’t quite agree with you on the survey weights. If NES oversampled blacks, I think a better way to do this is to run separate analyses for blacks and whites. In which case the weights for race won’t come in.

    Actually, given the small sample size for nonwhites, probably all you can do is fit your model to whites only, anyhow.

  5. John Sides March 10, 2009 at 8:44 pm #

    Andy, I thought of you when I made the plot. You’re right that the weights aren’t necessary to correct for the over-sample of blacks. This was sort of a brain freeze on my part: I’m analyzing only whites, after all. But I assume the weights are necessary to deal with other deviations from representativeness among the white respondents themselves.

  6. Andrew March 10, 2009 at 8:48 pm #

    Yeah, including the weights is a safe move; I just thought you’d overemphasized that point slightly.

  7. jodie March 12, 2009 at 5:53 pm #

    i like wat you have done