There’s no question that special elections are telling us that this year is a much better environment for Democrats than we’ve had in quite some time. In the graphic at the top of this post, you can see that Democrats have put up better margins (above the diagonal lines) than both Hillary Clinton (left) and Barack Obama (right) in a majority of cases.
But there’s a wealth of additional information lurking in the results—and it’s important to distinguish between what’s real and what’s just an artifact that vanishes on further scrutiny.
For instance, in special elections this April and May, Democrats outperformed Hillary Clinton’s margin by just 4 percentage points, compared to a whopping 24 points in January, February, and March. Does this mean the Democratic advantage is evaporating?
And Democrats may have outperformed Hillary Clinton’s 2016 margins by gaudy margins in many special elections, but the best overperformances were in low turnout affairs. Does this mean we won’t see Democratic overperformance in higher-turnout midterms?
Finally, we’ve had two regularly scheduled governor’s races in New Jersey and Virginia, with margins very close to Clinton’s. Does this mean special elections are just too “special” to mean anything?
The answers, it turns out, are “no,” “no,” and “no.” But it takes some digging to see why this is so. Join us as we dig together.
In our previous post on the topic of special elections, we explained that there are two variables that explain more than half of the variance in special election margin shifts. These variables are the shift in presidential margins from 2012 to 2016 within the district, and the 2016 presidential margin in the state the district is in. Together, they show Democrats are overperforming Clinton the most in districts in red states, especially districts that grew redder in 2016 at the presidential level. On the other hand, in blue states, and in districts that trended blue, Democrats are overperforming by smaller margins, or even underperforming.
Based on this, we wouldn’t expect to see large shifts in the margin in the statewide races in Virginia and New Jersey, or in some other notably high-profile races such as Georgia’s 6th congressional district. It also turns out that that these two variables can explain any apparent relationship between turnout and Democratic performance. What’s more, they also account for any apparent trends in Democratic overperformance over time.
First up: a detailed look at turnout and Democratic overperformance.
Turnout
The idea that Democratic overperformance disappears when turnout increases to midterm levels is understandably tempting. It is true that the races with the highest Democratic overperformance compared to 2016 have low levels of turnout. And it is also true that the highest-turnout special election—again, Georgia’s 6th—does not show overperformance compared to 2016. Not only that, but the high-turnout November elections for governor in Virginia and New Jersey also showed little overperformance compared to Clinton, appearing to bolster the case. (We’ll tackle them later.)
These two factoids make for an easy case of mental connect-the-dots. Here, the top 10 biggest Democratic overperformances are shown above the horizontal line, with the Georgia 6th underperformance below:
But it is a case based on cherry-picking. Because, as it happens, some of the specials with the worst Democratic underperformance are also found when there was low turnout.
Nonetheless, a plot of the full dataset does still show what appears to be a weak trend in support of the “higher turnout = weaker Democratic performance” argument:
But while there appears to be a relationship to the naked eye, the correlation between turnout and performance implied by the slope of the trendline is not meaningful in any statistical sense. (For a time, this relationship had looked meaningful until a large batch of special elections in New York in April.)
But—and this is the heart of the matter—even the appearance of a relationship between turnout and Democratic overperformance disappears after accounting for the lean of the state and the presidential margin shift within the district:
Now that trend has disappeared—almost literally. You can just barely make out the dashed line showing the trend, which is so flat that it pretty much overlaps with the X-axis. Instead, what’s obvious now is that low turnout elections are just as likely to feature both Democratic over- and under-performances — and more likely to feature either than high turnout elections. And that means that for high turnout elections, the lean of the state and the presidential margin shift account for even more of the variance than for the dataset as a whole.
We have previously shown that, using all of the special elections in 2014 and another pair of 2017 special elections in Georgia, if you accept the apparent correlation between turnout and performance and project forward to midterm-level turnout, your projections would be very, very wrong. Together with the graph above, this means that while we may consistently see an apparent (but not necessarily real) trend of decreasing Democratic overperformance with increasing turnout, increased turnout is not causing decreased performance.
The statewide elections
But if increased turnout doesn’t cause decreased Democratic performance, why is it that the Democrats in the governor’s races in Virginia and New Jersey only outperformed Clinton’s margin by a little bit?
As it happens, the Virginia and New Jersey governor’s margins were just about what we would expect if they had been special elections, based on the lean of the state and the presidential margin shift from 2012 to 2016. The same is true for last year’s Senate race in Alabama, and even the state Supreme Court race that took place in Wisconsin in April, despite the fact that it was officially nonpartisan.
In other words, the behavior of the major statewide races we’ve seen this year is entirely consistent with what we’ve seen in congressional and legislative special elections. There’s no need to wonder whether special elections or governor’s races are more predictive of November results, because both are telling us the same thing.
But don’t get too excited about how well these statewide results are predicted; with only four datapoints, it’s almost certainly a coincidence. All we should read from this is that these statewide elections are not wildly off the mark. Predictions for Virginia and New Jersey legislative contests showed plenty of deviation from what was observed, but the average error was still only about half of what we’ve seen with the special elections, even though almost all of those seats had incumbents running.
It’s understandable you’d ask how the margin in the Alabama Senate race could have been anywhere close to predictable with a nutcase like Roy Moore running. But the polling margin was actually about the same for both Moore and Luther Strange against Doug Jones, and, generally speaking, Generic Republican barely led Generic Democrat. Even if we assume that Strange would have won by 10 points, that would still have been a huge improvement for Jones compared to the 2016 presidential margin, and consistent with special elections overall.
The trends
We’ve noticed that Democratic overperformance for the first months of 2018 was even greater, on average, than it was in 2017. Yet polling measures such as Trump’s approval rating and the congressional generic ballot have been at best steady and at worst improving for Republicans. How do we reconcile these two observations?
Here are the trends in Democratic overperformance in special elections:
We saw a dip last November, a surge in February-March of 2018, and then more recently a dramatic fall with multiple examples of Democratic underperformance.
But what if we account for our two by-now familiar variables, state partisan lean and district presidential margin shift?
Plotting the residual of the regression for these two variables against time shows us … essentially no trend. Again, this is, in effect, all of the variability in Democratic overperformance that cannot be accounted for by those two variables. Basically, all the apparent trends in the graph above can be accounted for by which sort of districts are holding elections when.
Summary
In the end, we have a few neat and tidy conclusions that can help us understand what is going on with special elections and what they can tell us.
Statewide elections and special elections all behave in a consistent manner and are telling us the same story. We’re seeing shifts towards Democrats concentrated in districts that shifted red in 2016 and are in red states. Where we’ve seen smaller shifts towards Democrats, or even shifts toward Republicans, it is not increased turnout that is the cause. And the magnitude of the shift—overall, nationwide—has not changed over the past year.
