Addendum on Exit Polls
The findings presented in this analysis are derived primarily from New York Times exit-polls from the 2008, 2010, 2012, 2014 and 2016 national elections. In anticipation of questions or objections from readers, allow me to briefly elaborate why I chose to rely on exit-polling for this analysis, and why readers should trust (and act on!) the findings.
Are exit-polls reliable?
The main purpose of exit-polling is to help media organizations 1) determine the winner of presidential elections as soon as possible and 2) gather additional information about who voted and why, for the sake of framing how particular electoral outcomes came about. While early results of exit-polls can be misleading—especially in the hands of irresponsible or sensationalizing bloggers and journalists–-finalized exit-polls tend to match closely the eventual outcomes of most races, making them reliable enough for our purposes.
Don’t exit-polls seem to systematically over-represent certain groups or districts?
Of course, exit-polls by their nature focus on Election Day turnout, and they cannot easily incorporate early-voting results. However, this is not a problem for us, as early-voting patterns have not proven terribly predictive or decisive in terms of how races ultimately turn out.
Additionally, because exit-polls primarily serve to determine the outcome of races as efficiently as possible, they tend to oversample highly competitive states and districts. This is actually a good thing for our purposes—relative to swing states and districts, the prevailing sentiment or demographic composition of partisan safe zones don’t matter much in determining the outcomes of elections. And it is the predicted outcomes of (future) elections that we are concerned about here–so the bias towards competitive territories is just as useful for us as it is for the exit-pollsters.
Finally, for various reasons exit-polls also tend to oversample Democrats and Democratic-leaning constituencies. Again, for our purposes, this is not a problem. It means, if anything, our already-bleak narrative may be too generous to Democrats. It does not undercut our argument substantively—but instead renders our conclusions more urgent.
What about the margins of error?
It is true that the statistical margins of errors for exit-polls tend to be higher than other forms of polling, for reasons such as their strategic sampling bias. However, these concerns are largely theoretical (concerning the representativeness of the sample relative to the entire population).
In practice, exit-polls do tend to reliably track with the outcomes of races, and that’s what matters. Put another way, what is important to us is not so much the exact percentage of particular groups who lean one way or another in any particular contest, but instead, the longitudinal trend. To the extent that the New York Times’ methodology remains more-or-less consistent, we can be confident in the directionality of whatever they are capturing, while bracketing questions about representativeness.
When we see clear patterns that persist over the course of the last three presidential races and the last three midterms elections—and most importantly–when the story told from the exit-polls matches and seems to well-explain the actual electoral outcomes we’ve been consistently observing over the past decade, then we have pretty compelling reasons to trust the overall picture provided by the exit-polls. And indeed, that has been the case here.
But how pronounced are the trends, really?
In evaluating longitudinal trends one could choose to look for patterns that persist across all general elections regardless of whether they were presidential years or not–-for instance lumping together 2006, 2008, 2010, 2012, 2014 and 2016 into one dataset. On this model, one would still see a clear downward trend, but it would seem kind of like a roller coaster.
Consider the Democrats’ share of the youth vote over the last six cycles: 61, 66, 58, 60, 55, 55. The story is still a significant net loss over the last decade, but it could look like the numbers have been rising and falling, and perhaps one might expect the next election will look better than the last. However, such optimism would be unwarranted:
In reality, midterms and presidential races have drastically different demographic constitutions. For instance, 18-29 year olds amounted to around 18-19 percent of all voters during presidential years, but between 11-13 percent during midterm elections. And the subset of young voters who turn out during midterms tend to have different sensibilities than the broader and more diverse group who participate in presidential elections.
For instance, if we separate the Democrats’ share of the youth vote by presidential years v. midterms, a different picture emerges. First, the trend is unanimously down: in presidential years, from 66 percent to 60 percent to 55 percent; in midterms from 61 percent to 58 percent to 55 percent. So there is little reason to expect, from the trend, that the next race would show an improvement.
However, one also sees that Democrats generally perform better with young people in presidential years than in midterms–benefiting from left-leaning irregular voters drawn in by the race for the White House. Therefore it is noteworthy that the Democrats ultimately ended at 55 percent vote share for both the last midterm and the last presidential election:
The Democrats’ decline among young voters has been far more dramatic among the last three presidential cycles vs. the last three midterms (down 11 percentage-points from 2008 in the former vs. 6 percentage-points since 2006 in the latter). This convergence is very bad news; it means Democrats may not be able to rely on a presidential-year bump from the youth in 2020.
Similar patterns emerge across demographic categories. When you do apples-to-apples comparisons of midterms and presidential cycles, the trend is dramatically and unrelentingly negative for the Democrats and, to a lesser extent, positive for the Republicans (and third-parties).
Through regression, we can also see the aggregate trend and project through 2018 and 2020—as we did in the main essay for each demographic category. The results are unanimously bad.
A note on strategy
Trump and the Republicans are now the default in American politics. They are in power, and will likely remain in power–even if they are widely disliked–unless and until the public sees a better alternative (as opposed to a possible lesser evil). After all, Trump was elected with record-high unfavorable ratings to begin with!
This essay highlights apparent trends in the way Americans are voting, and suggests that if these trends hold into the future, the outlook is not so great for Democrats. However, as the 2016 election cycle should have powerfully driven home, and as I underscored in the conclusion of the article itself, there are limits to polls and predictions.
Ultimately, Democrats have to choose what they’re going to believe. It seems clear what would be more pleasant to assume (i.e. the Emerging Democratic Majority is on track, Trump is an ill-fated anomaly, etc.)—however, this is perhaps not the most useful proposition to work from.
Consider a variation on Pascal’s Wager: if the analysis presented here is somehow overly-pessimistic (again, the typical bias from exit-polls suggests the opposite), but Democrats buy into the argument nonetheless—what would happen? They would urgently invest time and resources, reworking their platform, outreach and messaging in order to shore-up white voters while simultaneously energizing and mobilizing minorities. That is, the only possible outcomes are positive.
On the other hand, if the analysis presented here is correct, but Democrats fail to heed the warning—instead perhaps doubling-down on their current strategies—the results would be catastrophic for them.
That is, strategically speaking, there is really no reason not to take these findings seriously, and a number of reasons why Democrats should. But we all make our choices…
Addendum on Projections
In the section of the essay where I discuss Democrats’ electoral trends along gender lines I make the following claim:
For those who are interested, here’s how I did that math:
- Through linear regression we can determine the slope and y-intercept of the lines “women” (m = -0.2286, b = 513.32) and “men” (m= -1.0571, b= 2171.2476).
- Using the slope intercept formula [y=m(x)+b], if we set our year (x) to 2020, we can solve for our predicted vote share (y) for women [0.51548] and men [0.359056].
- Using the data from Figure 2, we can determine the average electoral share of men (0.473333) and women (0.526667) during presidential cycles.
- Taking the product of how much each gender typically represents of the electorate and Democrats’ predicted performance within each group, we can determine how much of the total electorate each gender is expected to deliver for Democrats in 2020: Women 27% (0.51548 * 0.52667), men 17% (0.359056 * 0.473333).
- For our purposes (commensurate with most polling), men and women are considered mutually-exhaustive categories of the electorate. Therefore the sum of Democrats’ predicted vote share for men and women (0.44144 or 44%) is equivalent to the party’s total anticipated popular vote share in 2020–that is, unless they can somehow reverse the trend of the last decade (2006-2016).