KEY POINTS FROM THIS ARTICLE
— A surprising result in two special House elections last year, when Democrats ended up coming closer to winning an overlooked race in South Carolina than a nationally-watched contest in Georgia, showed an overreliance on past presidential election history to project House results.
— When assessing the true Democratic targets on this year’s House map, a more holistic approach is required, including looking at past down-ballot performance and other factors.
— Third-party voters, often overlooked, could be immensely important in determining the next House majority.
Modeling House elections
June 20, 2017 wasn’t like any other day, because it was an Election Day, hyped to a frenzy in the political world. Both national parties, and politicos everywhere, were on the edge of their seats waiting for results to come in. Polls would close soon in the extremely closely watched GA-6 special congressional election, as both sides hoped that each of their tens of millions of dollars invested in this race would not go to waste.
Personally, I was nervous for another reason, as I had put on the line a new and untested House model that was giving me numbers outside of both mainstream thought or even what the polling was showing. Where it looked on the surface like a virtual tossup with the possibility of Jon Ossoff (D) flipping the seat in Georgia, another special election in SC-5 would be decided the same night, although it was a seemingly safe red district. In fact, SC-5 became the butt of a lot of jokes on among election personalities compared to GA-6. However, I was staring at, and betting on, a very different outcome.
After performing well in the relatively predictable at-large special election in Montana and decently in the KS-4 special, my House model was trying out its first live test and was showing a shocking prediction: It showed Karen Handel (R) winning by three points in GA-6 with 51.49%, but it also showed a better Democratic performance in the overlooked South Carolina race. According to the model, Ralph Norman (R) was only projected to get 48.35% and lose to unheralded Archie Parnell (D). At the end of the night, Handel received 51.8%, winning by about three and half percentage points, while Norman took 51% of the vote, winning by a smaller margin than Handel (three points). There was still some error, especially in SC-5 as I will outline later, but the results bared out well, with a closer matchup in the overlooked seat that prognosticators called a Likely Republican race, while the Toss-up actually wasn’t quite as close.
In this piece, I hope to outline the couple of significant choices I’ve made in my House model that stand in contrast to other sites and election handicappers, and show a behind-the-curtain peek at some opportunities to improve the way we judge a race.
Why were Republicans favored in GA-6?
Before I go through why I believe that GA-6 was not a guaranteed Toss-up seat, it is relevant to take a detour and look to see how it became a marquee election in the first place. The race did not really engage until late January, when left-wing grassroots forum and data site Daily Kos put out a fundraising appeal to its list, outlining why the site believed the race could be competitive, and it went viral. The Daily Kos appeal admitted in general terms that voters in GA-6 typically vote Republican down-ballot, but it noted Donald Trump’s large drop from Mitt Romney’s 2012 vote share in the district as a sign that a Democrat could compete there.
This speaks to an overreliance on the Partisan Voting Index, an easy and readily available statistic to find that looks at only the presidential margin, and takes the average of the last two performances there compared to the national party performance. This makes PVI a poor measure of congressional performance when taken alone because it doesn’t take down-ballot performance — including, most importantly, congressional results themselves — into account. One of the principle differences in my model is that it uses the 2016 presidential results as a single benchmark, rather than the be-all, end-all in predictions, and it punishes or rewards members of Congress that establish an independent identity, which I will detail below. My model attempts to produce more accurate congressional predictions by separating a district’s presidential vote from congressional performance and by looking at raw vote instead of margin.
A first look to evaluate PVI is to show that it regularly underestimates the down-ballot effect on a partisan congressional seat, and it goes beyond missing an incumbent advantage, but also a partisan advantage. For Chart 1 below, I took the 2012 and 2016 PVI from their respective election cycles, and derived what the predicted vote share for GOP candidates would be from the GOP presidential vote that year (x-axis). I then plotted them against the y-axis, which shows what percentage the actual candidate received if their race was contested. The diagonal line is a reference line, and any point above that line was a Republican overperformance while any point below was a Democratic overperformance. The color is based on which party held the seat, and the year that data point was taken from. The yellow O-2012 (the O is for “open”) were seats that were changed post-redistricting, but had no incumbent running, and serve as a useful baseline that should not lean one way or the other.
Chart 1: Accuracy of pure PVI-based predictions
Note: Click on chart to see larger version.
Source: Cook Political Report
There are a couple of quick takeaways here. PVI underestimated the Republican share in Republican-held seats 74.0% of the time, but overestimated the Republican share in 85.7% of Democrat-held districts. You can also see this trend exists in both 2012 and 2016, even though 2016 featured a considerably higher share of third-party presidential voters than 2012 (about 6% in 2016 versus less than 2% in 2012), showing a bit of a robust consistency. So, is this because of an incumbency advantage or is there a partisan lean in the districts that PVI is missing? To explore that option, the boxplot in Chart 2 breaks the House elections up by year and incumbency status, with the error on the y-axis.
Chart 2: PVI error and partisan miss
Note: Click on chart to see larger version.
Side by side, you can find the quantitative error of PVI, where PVI understates the Republican share in GOP-held seats by three or four points, while it overstates the GOP share in Democratic-held seats by about four to five points. With the open 2012 seats registering as slightly Democratic, let’s say that PVI is typically off by about four whole percentage points, which is an eight-point margin! This is how some incumbents are regularly able to survive in crossover seats — districts where the party that won the House election differs from the party that carried the district at the presidential level — particularly when the national environment is neutral as opposed to a wave. This also does not change with open seats by a statistically significant amount, and not in the same direction in both cycles, so incumbency is not causing the error.
So PVI — or in essence, the presidential margin — already has a congressional partisan underperformance built into it, but the PVI in GA-6 was especially deceptive. This district voted heavily Republican down-ballot, and former Rep. Tom Price (R), who held the seat prior to serving a short stint as secretary of Health and Human Services in the Trump administration, had separated his identity from the top of the ticket the year before. Mitt Romney received 60.96% of votes in GA-6 in 2012, and Gov. Nathan Deal (R-GA) won 60.6% in his 2014 reelection. In 2016, however, Trump plummeted to only 48.3% of the vote, winning the district by only about 1.5 points over Hillary Clinton. Price took 61.7% of the vote while sharing the ballot with Trump after receiving 64.5% and 66.0% in 2012 and 2014, respectively. My model takes note of this pattern, with down-ballot Republicans performing significantly better in the district than Trump, and it also gave Republicans in the district a large bonus because of the recent ability to not get dragged down by Trump’s decline in 2016. When compared to the 2016 congressional vote, the GA-6 special election would produce a swing in line with other expected US House seats when shifting the generic ballot down but in comparison to the presidential vote, Ossoff lost the district by a little bit more than Clinton did (a 3.6-point defeat for Ossoff versus 1.5 for Clinton).
There was a large chunk of voters that was the cause behind Trump only winning 48.3% of the vote in GA-6 but allowed Price to take 61.7% as opposed to his Democratic opponent winning 38.3%. These voters clearly did not like Trump and while some held their nose for Clinton, they clearly had Republican sentiments, and many voters went with third-party options at the presidential level. Nationwide, Libertarian Gary Johnson and independent Evan McMullin increased third-party vote share in larger suburban areas, and took almost exclusively from Republicans. In 2016, these voters stayed faithful down-ballot. However, there is evidence that these voters are fleeing in droves.
Chart 3 below plots the third-party relative performance compared to past years on the x-axis and on the y-axis the change in the margin from Romney to Trump. The three counties that are in GA-6 are highlighted in red, and the size is by the number of voters in 2016.
Chart 3: Third-party performance impact on the margin change
Note: Click on chart to see larger version.
From this, you can see that there is a relationship between the two variables. As the third-party vote share increased in 2016 compared to past years, the presidential margin shifted towards the Democrats (p<.0001). Additionally, these three counties shifted to Clinton, and Trump’s numbers dropped, but the cause is mostly concentrated on third-party voters that also backed Tom Price. Most were bound to shift down-ballot as the generic ballot got worse, but it shows that we can reasonably expect those voters to shift first over those who backed Trump, which is the assumption made looking only at two-party share.
We can see the same thing in Virginia’s 2017 gubernatorial election, where now-Gov. Ralph Northam (D) won 53.9%-45.0% after Clinton won the state 49.8%-44.4%, building on Clinton’s margin by a few points. Down-ballot, the Democrats netted 15 seats in the state’s House of Delegates, all but one of which was carried by Clinton (Trump took one by about half a point). Up-ballot, the shift is marked by another strong relationship (p<.0001) when plotting Virginia localities by third-party performance in the 2016 presidential election and the shift in the Democratic margin between the 2013 and 2017 gubernatorial races, as shown in Chart 4.
Chart 4: The 2016 third-party vote impact on the 2017 Virginia gubernatorial race
Note: Click on chart to see larger version.
Here, 2016 third-party performance was extremely predictive of the swing in the populated counties, and of the rural counties becoming even more Republican leaning. Looking at the raw vote instead of the two-party vote share makes analysis much more accurate and gives a solid foundation of analysis for the upcoming 2018 election.
When trying to evaluate the competitiveness of a seat, it is a mistake not to disentangle the House candidate from the president and to compare seats where third-party support was higher than usual.
Why was the SC-5 race so close?
While the Georgia race focused too much on the 2016 presidential result in the district and didn’t separate what happens up and down the ballot, analysis of SC-5 made the exact opposite mistake. The presidential and top-line numbers had not changed much, starting with Romney winning 55.6% in the 2012 presidential race, then Nikki Haley taking 56.6% in the 2014 gubernatorial race, and finally Trump winning 57.3%, showing a steady but very slow Republican growth. Then-Rep. Mick Mulvaney (R, SC-5) received 55.5% in 2012, then 58.9% in 2014, and lastly 59.1% in 2016. The fact that there wasn’t a huge GOP overperformance and that the trend followed closely to the top of the ticket punished the Republicans in this race. Not separating from the top of the ticket typically doesn’t do much, but in a wave year where the top of the ticket is unpopular, we would expect the congressperson to plunge hard as well, and the model slaps a huge penalty in this environment.
Chart 5: Did Trump have a down-ballot effect?
The plot in Chart 5 shows the change from Romney to Trump on the x-axis and then the change in the GOP congressional candidate’s vote share from the 2012 to the 2016 cycle on the y-axis. The red line is a 1:1 ratio diagonal reference line. The SC-5 data point is on track where you would expect, but GA-6 shows a huge congressional Republican overperformance relative to the change from Romney to Trump. The other key difference is that third-party vote share in SC-5 was lower, so there was not as large a group of anti-Trump votes to split off and vote down-ballot for Democrats. If there was any effect, it would be tied entirely to the top of the ticket and how Trump’s approval and the generic ballot looked.
Map 1 below plots the relative third-party performance in 2016 from orange to green with the GA-6 and SC-5 district lines outlined in black. The map compares the total third-party vote share in each county to the national third-party share. It then looks at that localized performance compared to how well the third parties did in past elections in order to see if there was a relative over- or underperformance (green or orange, respectively) not only when looking at the totals, but when set up against past behavior, so that low third-party groups are not always punished, and high areas are not consistently seen as more favorable every cycle. SC-5 had lower third-party voting compared to GA-6, so those voters are not to blame for the close election.
Map 1: Third-party performance relative to historical performance in Georgia and South Carolina
As I mentioned above, my model was off, predicting a Democratic flip in SC-5. That result was driven by a factor that brings in a prediction based off a demographic regression that leans Democratic because the district it is slightly over 28% African American, and in a low-turnout special, it is likely that this demographic group didn’t really turn out in force. Democratic challenger Archie Parnell was aware of this, campaigning heavily to make inroads with this community with some limited outside help of only $275,000 for get-out-the-vote efforts, a fraction of the money spent to boost Ossoff in GA-6. With more investment and time in the African-American community, Parnell may have won. Instead, the Democrat underperformed the model by three points, though the media narrative turned Parnell into a strategic genius.
There should be a couple of takeaways from this piece. The first should be to rarely rely on the popular assumptions around a race, but instead to look at all the surrounding factors around a seat rather than focusing on one specific statistic or angle, like presidential margin. The second is that what applies to one seat doesn’t necessarily apply to all the seats, and it helps to separate the contests as much as possible and avoid easy catch-alls to explain away an election result. Each race is complex, and it makes sense to dig into the past for context behind the seat’s movements. It pays to look at as many past results as possible, and the 2016 presidential should be a single factor among many, not the defining factor.
|Noah Rudnick is currently a student at the Ohio State University and author of a column on Decision Desk HQ where he breaks down a new 2018 House seat every week. For more political commentary and to see his newest stories and model updates, follow him on Twitter @rudnicknoah|