The Democrats need to win 3 seats in the United States to gain a majority (they currently hold 48 and the Republican vice-president casts any tie-breaking vote). Their first pick-up opportunity comes tomorrow, in the Alabama special election to fill the seat of Attorney General Jeff Sessions.
Doug Jones has an unexpectedly high chance of winning the deep red state for the Democrats, with a recent polling average assigning Republican Roy Moore a barely significant 2.2 point lead. So how likely is it the Dems will pick up exactly 2 Senate seats in the 2018 midterms, thus needing to win tomorrow?
To do this I looked at every contested midterm and general Senate election since 1990. I excluded the (rare) cases when a third party candidate won a significant share of the vote and performed a simple regression of two party vote share on the president’s net approval rating (as a proxy for Congressional generic ballot — if you have that data please send it to me!), the state’s partisan lean, incumbency & whether the election was in the midterm or general cycle. I used these to predict each candidate’s two party share, and ran a probit on these predictions to forecast how likely each guy was to win.
How good is the model? Using data only from 1990–2010, it correctly predicted the winner of 92/101 Senate elections in the November 2012, 2014 & 2016 elections (there were three special election in 2014 & I excluded the 2012 Maine election & 2014 Kansas election due to the strong performance of third party candidates. The model correctly predicted the caucusing party of the winner of both those elections.)
The model predicted Scott Brown would defeat Elizabeth Warren in 2012 & Kelly Ayotte would defend her seat v. Maggie Hassan last year. Both were wrong, but only due to the model’s simplicity. If I gave senators who’d never won a regular election less of an incumbency advantage (a change that is probably justified), the model gives Warren the nod in 2012. And if I treat incumbent governors running for Senate as simply “incumbents” (surely justified), then the model sees 2016 for Hassan.
The model isn’t shy about calling winners (in 30 of the 101 races it gave on of the candidates a 98% or higher chance of winning the seat) but it assigned significant probabilities to each of the other seven underdog wins. (Cory Gardner 22%, Heidi Heitkamp 26%, Tammy Duckworth 31%, Thom Tillis 34%, Gary Peters 36%, Joe Donnelly 39%, Tim Kaine 49.8%). Even the simple four variable models works pretty well — candidates win when they’re supposed to, unless they’re a live underdog in which case they sometimes win but normally don’t.
If you know how popular each party is (or in our course, the President), whether it’s a midterm, who’s the incumbent and how blue the state is — there’s usually not much mystery left.
I assumed there were no incumbents in the Arizona, Tennessee & Minnesota special elections (Jeff Flake & Bob Corker have already announced their retirements. Tina Smith, Al Franken’s rumoured replacement is now apparently considering running in 2018 having initially demurred. Nonetheless I assumed it’s a vacant seat).
For 50,000 sims I assigned each incumbent a 5% chance of retiring with the exception, per PredictIt of Ted Cruz (6%), Joe Manchin (22%), Bob Menendez (37%), John Barroso (40%) & Dean Heller (55%).
To predict Trump’s approval rating I did a simple regression of a President’s net Gallup approval rating during midterms against his net approval in the middle of the previous December. No President has ever had a higher approval rating in midterms than 11 months prior, but the model actually predicts Trump’s net approval will tick up from -25 to -21: approval ratings consistently fall and revert to the mean — Trump’s unprecedented unpopularity make the latter outweigh the former. However, the prediction isn’t very certain — Trump’s net approval rating in each sim is draw from a Normal distribution with mean -21 and standard deviation 14.
Tim Kaine Isn’t Losing Virginia
We now have predictions for each of the 34 currently scheduled Senate races in 2018.
My predictions are generally more bullish on Dem prospects than those of expert prognosticators. Why? In a case like Florida, the experts are banking on Republican incumbent Governor Rick Scott challenging Democratic incumbent Senator Bill Nelson (something PredictIt thinks is very likely). Since my model doesn’t treats Governors the same as anyone else, it assigns Nelson near lock status. If I assume a Nelson-Scott matchup and treat the race as incumbent v. incumbent, Nelson is a 74% favourite — far from a lock and closer to the experts.
But in general, I suspect my predictions are more bullish on Dem prospects because … Cook, Roth & Sabato underestimate just how good an environment this is for Dems. Tim Kaine starts off well, as Virginia is bluer than average. The President, who is a Republican, is massively unpopular and likely to remain so which is big. And Kaine is running against that President’s party in the midterm, which is bigger. And Kaine is an incumbent, which is huge. Further, there are no “mitigating factors” to suggest the GOP has a shot. One month ago Democrats shellacked Republicans in Virginia. Kaine’s likely challenger is a racist who advocated Birtherism, three days ago
Kaine is popular and clean cut. Declaring Virginia “Likely D” as Cook, Roth & Sabato do is deeply pusillanimous. Roth, at least, is smoking something seriously funky to describe Tennessee as “Safe R”. An open seat, in a likely incredibly favourable midterm environment for Dems, is note “Safe R” even in a solid red state like Tennessee. That’s even before highly popular ex Governor of Tennessee Phil Bredesen announced his candidacy for the seat.
Similarly the three prognosticators are even less confident that Sherrod Brown wins reelection, assigning the state “likely D”, just a notch above “toss up”. Ohio is redder than Virginia, and assigning Brown less chance of reelection that Kaine is justified but still: Brown’s most likely challenger is Ohio State Treasurer Josh Mandel in a repeat of 2012. Brown won that race by 6% & in 2018 he’ll be running against a preternaturally unpopular President in the midterms, as opposed to under a mildly popular President in a general election year. Sherrod Brown is very likely to win reelection and merits more than a “likely D” rating.
On the other hand I suspect the model is slightly overconfident in the chances of Dems holding on to their “lock” seats. Republicans picked up Hawaii only 16 times in 50,000. I don’t expect Hawaii to turn red, but weird shit happens.
So what does this mean for control of the Senate?
If Doug Jones doesn’t win tomorrow, the model gives Democrats an 8.6% chance of winning control of the Senate in next year’s election.
If Doug Jones wins tomorrow, the model gives Democrats a 24% chance. That is, the model thinks the Dems pick up exactly 2 seats in next year’s elections about 15.4% of the time.
While Dems are favoured to pick up Arizona and Nevada and have good chances in Tennessee, the landmine navigation in their 26 incumbent is pretty tricky. They’re favoured to win each individual seat but even in a good environment they’re going to lose a Montana or Indiana or West Virginia pretty often. Even if they win Arizona, Nevada and Tennessee, they’re still only 36% to win three or more net seats.
The Political Climate Is Really Really Important
When Trump’s net approval rating (which as a reminder I’m using as a proxy for generic congressional polling) drops between -33 & -37 Dems are 15.4% to win more than 2 seats & 38.5% to win more than or exactly 2 seats. That is, if you think Trump’s approval rating will continue to slide, the Senate becomes close to a toss up if Jones wins tomorrow.
If, however, Trump’s net approval improves 10 points to between -13 & -17 are 3.8% to take the Senate in the “Jones loses” scenario and 14% if Jones wins. And If Trump improves to between -3 and +3 our model thinks the jig is up for Chuck Schumer. In this scenario Dems take the Senate just .9% of the time if Jones loses tomorrow and 4.7% if he wins.
The Dems have a chance — thanks to Donald Trump
While Dems need a *lot* to go right to take the Senate they have a chance, especially if Jones wins tomorrow. But they only have this chance but Trump & Republicans are so unpopular. Under a generic Republican president like Mitt Romney, the 2018 Senate map is a layup for the GOP. Incumbents are tough to beat and midterms are always tricky for the governing party but it’s a really, really favourable map & if Flake and Corker ran under Romney (which is likely) it’s hard to see Democrats winning those states (Tennessee would be close to a lock for the GOP).
I’ve already mentioned a couple of reasons why the model might overestimate Democrat winning chances. But there’s two big reasons it might underestimate the chances of Dems winning the Senate.
i) Defections. In scenarios where Democrats have exactly 50 Senate seats after the 2018 midterms, they’ll likely hold a majority in the House. And “moderate Republicans” might choose to switch parties in order to have influence for their states. There aren’t many candidates to cross the house it’s plausible Lisa Murkowski & Susan Collins could go blue.
ii) A second pickup in Arizona. Senator John McCain’s health is not great & the model makes Democrats a -200 favourite in a vacant Arizona Senate election. Dems should also recruit a strong candidate for Arizona’s governorship — if McCain’s seat isn’t up for grabs next November the chances of a vacancy arising in the next two years are pretty good.