In prior analyses (here, focusing on 2018, and here, considering trends 2010-2018), we asked a key question about Democratic strategy in 2020: should the focus be on turning out likely Dem voters, or changing the minds of folks who are less likely to vote for Team Blue? (TL;DR: We concluded that we should definitely get our base to turn out, but we shouldn’t neglect subgroups outside of that group that align with us on key issues).
Here, we ask an obvious follow-up question about turnout: if we want to mobilize likely Democratic voters, on whom should we focus? Is that answer the same in every state or district? There is a tendency to imagine that the voting preferences of a demographic group are the same everywhere. That, for example, young voters in Texas are the same as young voters in Michigan. But looking only at national averages obscures lots of interesting regional variation. That regional variation matters, since house and senate campaigns contend for voters in specific places and, as long as the path to the presidency goes through the electoral college, the presidential race also has a strong geographic component.
In this post, we focus specifically on college-educated voters, and mostly on female college-educated voters. We were inspired to begin our analysis here by one of our favorite election modelers, Professor Rachel Bitecofer, who generously spent some time chatting with us last month about our related interests. She reminded us of a core idea in her spot-on analysis of the 2018 house races: the presence of “large pools of untapped Democratic voters,” including college-educated voters, makes for places where Democrats can (and did in 2018!) outperform compared to 2016.
In our work, we always try to keep an eye on how progressives and Democrats can best use their time and/or money in upcoming elections. Our values make us particularly interested in work that emphasizes registration and voter turnout. In this post we examine if and where a college-educated-voter turnout drive might be most effective.
In our research pieces so far, we’ve looked only at aggregate data, that is data which comes from adding together a large number of people: census counts or election results, for example. In this piece we look at some per-person data, namely the Cooperative Congressional Election Study, or CCES, which surveys about 60,000 people in every election cycle. For each survey response, the CCES includes geographic and demographic information along with opinion about various political questions, whether the person is registered to vote, and whether they voted and for whom in elections for Governor, House, Senate and President, whenever each is applicable.
The geographic information allows us to start estimating a variety of things at the state level, something that isn’t possible using only aggregate data. We do this using Multi-level Regression (the “MR” of “MRP”, which stands for Multi-level Regression with Post-stratification), a technique explained in more detail here and about which we will have an explainer in the coming months. Very briefly: this method allows you to use all the data (e.g., all female-college-educated voters over 45 who voted for a Democrat or Republican) to get a baseline and then use the data in a specific “bucket” (e.g., those same voters, but just in Texas) to make an improved inference for that set of people. In other words, it balances the abundance of the non-specific data with the sparser local information for improved insight.
If we are hoping to boost turnout among Democratic leaning voters in battleground states or crucial races, it helps to know which voters are most likely to be Democratic leaning in that particular state or district. This data and these techniques allow us to do just that.
(Quick aside for data geeks: Our specific model uses a survey-weighted, logistic MR to fit a binomial distribution to the likelihood of a Democratic vote in the 2016 presidential election, among voters who voted D or R, in each of 408 = 51 x 2 x 2 x 2 groups: (states + DC) x (Female or Male) x (Non-College Grad or College Grad) x (Under 45 or 45 and over). Also, as we’ve said in earlier posts, we recognize that these categories are vast oversimplifications of the electorate. We are limited by what information is present in the survey—Sex is limited to “Female” and “Male” in the CCES—and by computational complexity, which is why we’ve limited our education and age breakdown to two categories each. Its also important to note that the accuracy of these inferences depends on the number of voters in each category. So the estimates for more populous states are likely to be more accurate.)
In this post, we analyze the 2016 presidential election to measure Democratic voter preference among college-educated voters in each state (and DC). Given the results in the house races in 2018, we think that most college-educated voters have become more likely to vote for Democrats since 2016. But there are inevitable differences between presidential elections and house races and we want to start with a straightforward question: If the 2020 electorate was like the 2016 electorate, where might a college-educated-voter turnout drive be useful?
Before getting into the results, let’s introduce a metric that we call “Votes Per Voter” (VPV), which reflects the “value” to Democrats of getting a single person in a particular group to show up on election day.
Here’s how VPV works: Let’s say that our model tells us that in a certain state, young, female, college-educated people vote Democratic about 72% of the time. If we increased turnout among them by 100,000 voters for 2020 and they were exactly as likely to vote Democratic as in 2016, how many votes would that extra 100,000 voters net the Democratic candidate? Of that 100,000, 72,000 (72% x 100,000) would vote for the democrat and 28,000 ((100% - 72%) x 100,000) for the Republican. So Dems would net 72,000 - 28,000 = 44,000 votes. In general, if \(x\%\) of voters will vote for the Democrat, each such voter is “worth” \(2x-100\%\) votes. In other words, the VPV for this group is 0.44 (44,000 net Democratic votes out of 100,000 new voters in the group who show up to cast ballots). Note that a group with a Democratic voter preference of 50% has a VPV of 0, a group that votes Democratic 60% of the time has a VPV of 0.2. And a group which votes for Democrats less than 50% of the time has a negative VPV.
This metric is useful because it highlights the connection between voter preference and votes gained by additional turnout. A group which leans only slightly Democratic is not a great place to invest resources on turnout since each voter is only slightly more likely to vote blue. This is reflected by the near 0 VPV.
As a side note, this is why changing people’s minds can seem a more appealing avenue to getting votes: changing a Republican vote to a Democratic vote has a VPV of 2 (200%): one vote is lost by the Republican and one is gained by the Democrat, so each such voter is “worth” 2 votes. (More on that in a later analysis—stay tuned!)
Let’s look first at female college-educated voters and how their VPV varies by state and by age. This demographic, one that has been trending strongly Democratic since the early 2000s, is particularly important for Democrats. In the chart below we show the VPV of female college-educated voters, split by age at 45, for each state (and DC and the Nation as a whole). We’ve ordered the states (plus DC and the nation as a whole) by increasing VPV of female college-educated voters under 45.
We also looked specifically at some states we think are worth highlighting, including several classic presidential “battlegrounds” and others (like Texas and Georgia) that are moving towards battleground status and have important house seats in play. we’ve added male voters here as well, in order to clarify the focus on female college-educated voters as a good source of Democratic votes:
There’s a lot of detail here, but we’d like to make a few main observations:
This early in the cycle it can be difficult to find organizations focused on specific groups within states. But there are some groups which are clearly working in the same vein. If you know of local organizations doing work on youth/college-educated voter turnout, please email us with that information and we’ll update this post.
Want to read more from Blue Ripple? Visit our website, sign up for email updates, and follow us on Twitter and FaceBook. Folks interested in our data and modeling efforts should also check out our Github page.