Analyses of the public’s political opinions from Pew Research (and to an extent Gallup as well) often follow a certain script. On whatever topic they’re covering, they’ll usually start with the current numbers. Then they’ll compare the latest survey with older ones to look for trends over time.

Then they’ll break things down by party affiliation and/or liberal-conservative self-identification. For most kinds of issues, these breakdowns show very large differences—self-labeled liberal Democrats really tend to hold liberal views on various issues while self-labeled conservative Republicans really tend to hold conservative views. These kinds of splits are interesting, but often carry tremendous causal ambiguity. They result from complex mixtures of people who hold specific issue opinions because of their party or ideology along with people who favor a party or ideology because they have a given set of issue opinions.

To give a better sense of what’s driving things, the last step in the analysis often shows various demographic splits. In recent Pew political posts, for example, they showed trust in government broken down by gender, age, and education, views on the rich and the poor broken down by gender, education, and income, and budget preferences broken down by gender, age, education, and income.

Overall, it’s a pretty good script. But I’m frequently puzzled by one aspect: the choice of demographics. Gender, age, education, and income often contribute to political divisions. Yet they’re mere ripples in the pool these days compared with the crashing waves of religion and race.

In this post, then, I want to show as plainly as I can that the contributions of religion and race to political differences are typically much larger than the contributions of other standard demographic categories. That is, if you’re doing a demographic breakdown of public opinion on political topics, you’d usually want to start with religion and/or race, because they’re the biggest deals. In addition, I want to show where religion tends to be especially dominant versus where race tends to be especially dominant.

And that’s the chart below. Basically, I took a few years of recent Pew data, grouped a number of their individual questions into multi-item opinion measures (on rich-poor economic redistribution, on homosexuality, etc.), and used multiple regressions to predict differences in those political items as a function of (1) information on religious identity (whether folks are evangelical, Catholic, atheist, etc.) along with frequency of religious service attendance, (2) racial and ethnic information along with immigrant status, (3) education, (4) age, (5) gender, (6) family income, and (7) region of the country (South, Northeast, etc.) along with population density (urban/suburban/rural).

You can see in the chart that religion/church (the black bars) and race/immigrant (the green bars) are just way bigger deals than education, age, gender, income, and region/density. Further, there are some kinds of items where race/immigrant variables are particularly big deals (party identification along with views on rich-poor issues, immigration, gun regulation, racial issues, and white nationalism, which combines views on immigration, race, etc.), while there are other kinds of items where religion/church variables are clearly the dominant demographic predictors (self-labelled liberal/conservative ideology along with views on homosexuality, abortion, marijuana legalization, environmental regulation, and Middle Eastern conflicts).

(Notes: The displayed values are additive contributions to the overall multiple R in forward stepwise OLS regressions. The predictor set included a large number of binary categories—being black, being Catholic, having a college degree, having an income below $40k, going to church at least one a week, and so on, and so on—as well as interactions involving all predictors. I entered individual predictors based on which had the biggest impact at any given step.)

The demographic categories other than religion and race are never seriously big deals (at least with this set of political items). Sometimes they’re moderate deals, though. Income is a moderate deal in predicting rich-poor positions. Education is a moderate deal in predicting views on immigration and guns. Age differences are a moderate deal when it comes to immigration and marijuana legalization. And so on.

Some of these demographic differences would be larger in a stand-alone analysis. Here, I’m using a regression analysis that accounts for the biggest differences first, and then shows only the marginal contributions for less important items. So, for example, age differences in issue opinions and partisanship are in part driven by the fact that younger groups are more racially diverse and less religious. Thus, when religion and race go into the models at earlier steps, the remaining marginal contributions of age differences aren’t typically very large. The story is similar with regional and urban/rural differences, which are largely driven by racial, religious, and educational differences.

Also, there are some key demographic items that aren’t typically measured in Pew samples. From other sources, for example, I’ve found political differences based on sexual orientation, veteran status, and occupational information.

Why avoid religion and race?

So it’s puzzling to me that Pew analyses often highlight age, gender, education, and income while often avoiding religion and race. I suspect the neglect of religion relates in part to the fact that Pew has separate groups focused on politics and religion, so perhaps the politics folks don’t like to crowd into the religion folks’ turf. I also suspect some of it relates to the complexity of the religious divisions. To really find the religious fault lines, you have to spend some time grouping and regrouping the categories. In my analyses, I’ve usually ended up settling into a not-entirely-obvious system that combines Mormons and non-Catholic evangelicals into one category, other Christians in another category, “nothing in particular” and those with missing information in another, specific non-Christians (Jews, Buddhists, etc.) in another, and then atheists and agnostics in yet another. There’s no great a priori insight that drives any of this for me; it just tends to be a set of categories that carves the sample effectively when looking at political differences.

I also suspect that lots of people just don’t really enjoy thinking about religious and racial differences, and particularly don’t enjoy noticing just how much of our current political differences are attributable to these sources. There’s something creepy about it—it’s too close to home and it’s too near the bone, and all that. Or maybe that’s not it; I really don’t know.

There seem to be at least three things that were true of Millennials in last year’s presidential election. First, they heavily favored Sanders over Clinton in the Democratic primary. Second, they heavily favored Clinton over Trump in the general election. But, third, they were substantially less likely than older folks to vote at all.

The charts below give a look at these patterns using data from the 2016 American National Election Studies (ANES). (It’s important to remember that this is just one sample. As I showed in a prior post, there are various differences among the ANES, the Cooperative Congressional Election Study (CCES), and the exit polls.)

If you’re used to thinking about exit polls, the big difference here is that I’m also showing non-voters. According to the ANES results, substantially more Millennials than older folks didn’t vote in the primaries (73% vs. 51%) and didn’t vote in the general election (38% vs. 20%). This isn’t something new—younger folks are usually quite a bit less likely to vote than older folks.

The ANES numbers, however, are almost certainly underestimating non-voters across the board. According to the United States Election Project (which uses actual vote tallies rather than after-the-fact surveys), around 41% of eligible voters didn’t vote in the general election. This is substantially higher than what the ANES sample suggests. In fact, to get to a 41% non-voting total, you’d have to assume something like a bit over half (rather than 38%) of Millennials and a bit over a third (rather than 20%) of older generations not voting. The problem with after-the-fact surveys is in part that some non-voters lie about voting, but it’s also that voluntary surveys disproportionately pick up the kinds of people who have opinions and don’t mind sharing them—that is, the kinds of people who are more likely to vote in the first place.

Millennials and the primaries

The top two charts show the primaries. And, sure enough, the ANES data suggests that, when they voted in the Democratic primaries, Millennials overwhelmingly chose Sanders over Clinton. But keep in mind that these data also suggest that older Democratic primary voters chose Clinton over Sanders in about the same overwhelming proportions. Here too, though, there are reasons not to oversell the exact numbers. The ANES sample gives Clinton a bigger total margin over Sanders (with about 59 Clinton votes for every 39 Sanders votes) than analyses based on the actual vote totals (where Clinton received 55 votes for every 43 Sanders votes). Also, the CCES sample shows Clinton running almost even with Sanders among Millennials, something that seems very unlikely given the ANES and exit poll results, but nonetheless represents a cautionary data point.

While Millennials seem to have heavily favored Sanders over Clinton in the primaries, their actual favorite option by far was to not turn out to vote (again, even the 73% non-voting number in the ANES sample for Millennials in the primaries is probably substantially too low). And, even among those Millennials who turned out, there were probably at least as many non-Sanders primary voters as Sanders voters. If you neglect these points, it’s easy to overstate Millennials’ support for Sanders.

Millennials and the general election

In the general election, we see again that Millennials were a lot less likely to vote than were older generations. And, as I discussed earlier, the ANES non-voting estimates for the general election are too low.

But for those who did vote, Millennials substantially preferred Clinton over Trump. Millennials also were more likely than older generations to support third-party candidates.

A big reason why Millennials generally favor Democrats over Republicans relates to generational differences in demographics such as race and religion. This shows up clearly in the CCES sample (which I analyzed in prior posts on Clinton/Trump voter demographics). Just looking at Clinton vs. Trump general-election voters in the CCES data, Clinton got 64% of the two-party vote among Millennials while she got only 48% of the two-party vote among older generations. That’s a 16-point gap.

And while a 16-point gap might seem like a big deal, it’s really not when you compare it to various bigger deals. So, for example, in the CCES data, there’s a 49-point gap between whites (42% voted for Clinton over Trump) and blacks (91% voted for Clinton over Trump), and there’s a 37-point gap between evangelicals (34% voted for Clinton over Trump) and non-Christians (71% voted for Clinton over Trump). Start combining such items—focusing, say, on white evangelicals—and the gaps grow even larger.

In fact, it turns out that the lion’s share of the Millennial gap in the CCES is due to the fact that, compared with older generations, Millennials have more racial minorities, fewer evangelicals and other Christians, more LGBT folks, and fewer military veterans. In short, what begins as a 16-point Millennial gap in the two-party 2016 vote gets reduced to a mere 5-point gap when statistically controlling for race, religion, sexual orientation, and veteran status.

While these fundamental demographics can explain most of the Millennial gap in the general election, they can’t, as far as I can tell, explain much of the Millennial gap in support for Clinton vs. Sanders. I have yet to see anything that really explains the strong generational splits within the Democratic primary (e.g., when I analyzed the issue positions of Millennials, it turned out that they’re actually not unusually liberal on Sanders-emphasized redistribution issues, even though they are unusually liberal in some other areas, such as views on homosexuality, marijuana, Middle Eastern conflicts, and immigration).

Another thing we don’t really know is the future. There are some safe bets, though. Like other generations before them, Millennial voter participation is likely to increase as they age. It’s also likely that, for the time being, given their demographics, Millennials will continue to prefer Democrats over Republicans when they do vote. Eventually, though, Millennials and those who come after them will inevitably force changes in the current party coalitions—there just won’t be enough white Christians around to support a viable national party organized primarily around white Christians, and so the parties will continue to evolve.