As the election approaches, we’re reminded that demographic features matter in predicting voting patterns. But which ones matter most and how should they be combined?
I looked at the U.S. General Social Survey from 2002 to 2012, limited to people who reported voting for either Democrats or Republicans in presidential elections (excluding non-voters and third-party voters). Here’s the simplest way I can think of to describe the results.
(1) African Americans tended to vote for Democrats regardless of other demographic features. Overall, Democrats have won about 92% of the two-party vote from African Americans, without any other features being especially important in changing that baseline percentage.
For the rest of the voting population, the predictors differ a bit for those with at least some college education versus those who never completed a year of college.
(2) For people who are not African American and who have not completed a year of college, here’s how to produce a pretty good estimate of their presidential votes:
- Give 1 point for each of the following that apply: (a) doesn’t go to church every week, (b) parents were immigrants, (c) Catholic, (d) not Christian (including people with no religious affiliation), (e) family income less than $75,000, (f) lesbian, gay, or bisexual, and (g) Latino, Asian, and other non-white race/ethnicity.
- Cap that number at 5 (i.e., if it’s higher than 5, bring it down to 5). Multiply it by 12.
- Add 19.
The result is the approximate percentage of the two-party vote going to Democrats. So, for example, for people who haven’t been to college and (1) are not frequent churchgoers, (2) have immigrant parents, (3) are Catholic, (4) have family incomes less than $75,000, and (5) are Latino, such folks voted for Democrats over Republicans around 79% of the time over the past several presidential elections (5 times 12 plus 19). On the other hand, for people who haven’t been to college and are richer, white, heterosexual, churchgoing Protestants with native-born parents, only around 19% voted for Democrats over Republicans (0 times 12 plus 19).
(3) For people who are not African American and who have completed at least a year of college, here’s how to produce their estimate:
- Give 2 points if they are not Christian (including people with no religious affiliation).
- Give an additional 1 point for each of the following that apply: (a) Latino, Asian, and other non-white race/ethnicity, (b) doesn’t go to church every week, (c) lesbian, gay, or bisexual, (d) never married, (e) graduate degree (MA, PhD, MD, JD, etc.), (f) female, and (g) Catholic.
- Cap that number at 6 (i.e., if it’s higher than 6, bring it down to 6). Multiply it by 13.
- Add 14.
Again, the result is the approximate percentage of the two-party vote going to Democrats. Former Congressman Barney Frank, for example, is Jewish (2 points), doesn’t go to synagogue every week (1 point), is gay (1 point), and has a graduate degree (1 point) — people with Frank’s demographic profile have voted for Democrats over Republicans around 79% of the time (5 times 13 plus 14). Congressman Paul Ryan, on the other hand, is Catholic, but otherwise doesn’t get any points — people with Ryan’s demographic profile have voted for Democrats over Republicans around 27% of the time (1 times 13 plus 14).
There are many ways to slice these pies. I’ve seen a lot of talk (maybe too much talk) recently about unmarried women, seniors, white evangelicals, and other categories that didn’t end up being particularly important in my models. (In the GSS data, I find that age effects are mostly covered by race and religion/church attendance, that with non-college-educated non-African Americans the effects of gender are mostly covered by income and the effects of marital status by income and religion/church attendance, that differences in religious fundamentalism are largely covered by education and church attendance, that the whiteness of white evangelicals isn’t more predictive than the whiteness of white Catholics or nonreligious whites, and so on.) But different samples can break in different ways.
Now, go forth and predict.