Veritium Political Insights
We Analyzed over 300 Senate Races Dating back to 2004: Here's What We Learned
Key Takeaway 1: The most recent presidential result in a state (relative to the national average) is the most consistent predictor of senate results
The most recent presidential result in a state (relative to the national popular vote) had the largest effect size and was the most significant out of any predictor of Senate results. For midterm elections, this variable is the presidential result from 2 years ago, and for Senate elections in presidential years, this variable is the result from 4 years ago. The presidential result has become more and more predictive throughout the past decade as Senate races have become more and more nationalized. The size of the effect has stayed steady from 2008-2018 and does not differ significantly depending on whether the Senate election occurs in a presidential or midterm year.
Key Takeaway 2: Democratic Senate candidates tend to slightly overperform expectations
Unsurprisingly, Democratic candidates tend to slightly overperform expectations in Senate races. Since the Senate map naturally strongly favors Republicans (analysts have estimated there would be a 90% chance the Republican Party controlled the Senate at any one time if the parties were equally extreme with regards to the median voter), Senate Democrats have formed more of a “big tent” caucus and Senate Republicans utilize their advantage to pursue an agenda well to the right to the agenda pursued by Republicans in other elected positions. In 2020, the median Republican was rated by GovTrack as a .83 on a scale of 0 (most liberal) to 1 (most conservative), while the median Democrat was rated .25, roughly 50% closer to the middle. Democrats such as Joe Manchin (D-WV) frequently break with party leadership and Manchin has experienced maintained electoral success even as incumbency advantages decrease. Our regression models predict that in a straight tossup race (meaning a perfectly neutral NPE, a perfectly even state, no incumbent, and no challenger that has held a major statewide office) a Democratic nominee would be favored by about 1 point and have roughly a 56% chance to win.
Key Takeaway 3: The incumbency advantage is severely decreasing over time
Our model fits the incumbency effect decrease over time, modeling a 4.2-point drop from 2008 to 2020 (16.2 to 12.1 points). Polarization has been steadily increasing since Barack Obama was elected president, and the campaign and rhetoric of Donald Trump throughout the mid-to-late-2010s accelerated this effect. With growing polarization on the presidential stage, Senate elections tended to become referendums on national party leaders. As the incumbency advantage was primarily fueled by split-party ticket voters, increased straight-ticket voting decreased this advantage and ultimately drove out multi-term incumbents, such as Claire McCaskill (D-MO) in 2016. Senate races in 2020 had the lowest incumbency advantage out of any year in our sample, despite it being a presidential year, at an average of only 9.8 points incumbency advantage per race. It is not unreasonable to expect a future where incumbency advantages are only in the 6 to 8 point range.
Key Takeaway 4: The incumbency advantage is significantly lower during midterms
The effect of incumbency is 5.3 points smaller in midterm years than in presidential election years. This difference is primarily driven by the fact the voting population in mid-term elections is generally more interested in politics, and therefore more partisan than the population that votes in the general election. Since the incumbency advantage is built on split-party voters, the more partisan electorates in midterms lessen the incumbency advantage. There are indications this effect may disappear, or at least lessen, as the midterm and presidential electorates become similar as a higher proportion of eligible voters vote in midterms. The adjusted difference between the 2016 and 2018 incumbency advantages was only 2.9 points, by far the smallest in the sample.