Special elections and the Trump Vote

By John D. Johnson and Charles Franklin

Special elections for the Kansas 4th and Georgia 6th Congressional districts have drawn tremendous attention as harbingers of Republican and Democratic fortunes under President Trump. The 6.8 point GOP victory in KS-4 on April 11 fell below most expectations for a district Trump won with 64.6 percent of the two-party vote in November. The Georgia 6th is viewed with even more anticipation as HHS Secretary Tom Price’s former seat and as an example of possibly shifting metropolitan districts. California 34, a heavily Democratic district where Trump received just 11.4 percent, also has held a special election this spring, though it drew far less attention.

A fundamental problem with focusing on these congressional seats is their small numbers which makes comparison and generalization to other districts difficult. We address this issue by exploiting the more numerous state legislative special elections since November 2016. To date there have been 16 of these (including “jungle primaries” in Louisiana and some other unusual circumstances — see the note at the bottom for what we include and how), a number which will grow to 37 by the end of August. Legislative elections are of interest in their own right given heavy Democratic losses in legislatures over the past 6 years and the role of legislative control in state policy making in general and particularly in upcoming redistricting following the 2020 census.

The simplest comparison is whether the Republican special election vote over- or under-performed Trump’s vote in the district in November. (We use percent of the two-party vote to exclude 3rd party votes. The results are substantively unchanged if we use percent of the total vote or vote margins.) The chart below shows all 18 special elections since the presidential election by date.

Setting aside the four one-party (“uncontested”) races, the GOP special election candidate has over-performed the Trump vote in 4 legislative races, underperformed in 8 legislative races and in the two special congressional races so far. The mean, including the two congressional elections is -2.2 and the median is -4.1. Excluding the congressional races the margin is smaller, a mean of -.95 and median of -2.7. By a 2–1 ratio Republican candidates are underperforming Trump, but the size of that margin is rather modest, especially in the state legislative races.

The two congressional races produce similar under-performance despite one district being heavily Democratic and the other heavily Republican. We can’t generalize from two cases but it is interesting that such different districts resulted in similar Republican under-performance compared to Trump’s vote (-11.2 in KS-4 and -7.9 is CA-34, based on two-party vote percentages).

It is entirely possible that legislative special elections are fundamentally different from congressional special elections, but the best way to know is to explicitly make the comparison of how Trump support in November relates to legislative and congressional elections in 2017. This is what we do next. See the data note below for details of how the data and models are specified. We exclude the 4 legislative elections that were not contested by both major parties (for brevity we call these “uncontested” though there might have been multiple candidates from a single party), and omit the congressional races from the fit estimated for the remaining 12 legislative elections.

The chart below shows how the Republican percentage in all the special elections so far has compared to Trump’s percentage in those districts. Congressional elections are shown in red and the 4 uncontested races are in gray. Both congressional and uncontested races are omitted from the linear fit, shown in blue. Results above the black diagonal line are races where the Republican candidate over-performed Trump and those below the diagonal are cases of under-performance.

While we saw a small average under-performance above, this is not large enough to reach statistical significance in the linear model. We cannot reject the null hypothesis that Republican performance in the special legislative elections is simply equal to Trump’s performance in the district. While not significant, the model does reflect the slight average underperformance: in districts where Trump received 50% of the vote, the model estimates a Republican share in special elections of 48.2 percent, a small 1.8 percentage point under-performance.

If legislative results perfectly mirrored Trump’s performance we would find an intercept of zero and a slope of 1.0 — a null hypothesis we fail to reject with p=.61. The data so far do not support the claim that GOP legislative candidates are doing significantly worse in special state legislative elections than would be predicted from Trump’s November results. Nor are they doing better.

There is considerable unexplained variability in these elections. Trump’s share of the November vote explains 59% of the variance in the GOP share in these legislative elections and the regression fit has a standard error of 13.6 points.

The conclusion then is that so far there is insufficient evidence to conclude that there is a Democratic surge across these 12 state legislative special elections.

Statistical tests of this sort on a small number of cases should be seen as adding a bit of discipline to our interpretations, and certainly not as definitive conclusions. The special election year is yet young and these results will change and become more convincing as we add more elections through the year. But the statistical tests caution us not to leap to conclusions based on a single race or even this set of races.

If we add the two congressional special elections so far to the model, the conclusions about legislative races are unchanged, but the estimated congressional coefficient is -9.8 points below the predicted fit due to Trump’s vote. This coefficient is not statistically significant, unsurprisingly given that it rests on just 2 cases, and so is not conclusive. It does suggest congressional Republicans may underperform Trump to a greater extent than do GOP state legislators. This raises two related questions. First, the direct connection between members of congress and Trump’s policy initiatives may lead voters to punish GOP congressional candidates as a result of currently low approval of Trump’s performance as president, and possibly more so than for state legislative candidates who are less directly tied to Trump’s actions. The verdict is clearly still out on this from a statistical point of view, though the negative coefficient is a first hint.

The second question is whether Democrats will improve their performance in legislative contests compared to their loss of over 900 legislative seats during the Obama administration. Historically there is good evidence that even as the president’s party typically loses seats in U.S. House midterm elections, so too does the president’s party suffer in state contests at the same time. During the Obama years, the consequences of these gains for Republicans at the state level have been substantial, and any reversals under Trump would be signaled by a drop-off from Trump’s support in these early legislative special contests. So far, we have little evidence of such a systematic decline.

As the special elections of 2017 play out in the coming months, this analysis will provide stronger empirical evidence concerning how the Trump presidency is affecting both legislative and congressional races. With 252 state legislative seats at stake in general and special elections through the rest of 2017, plus a handful of congressional special elections, we will have an early indicator of how support for Trump in 2016 may relate to the fortunes of Republican and Democratic candidates in the midterm.

Others have posted on this topic. See


Subscribers only:


Two related nice tweets:



DATA NOTE: We have included every special election in order to maximize cases. Some of these present special circumstances compared to the usual two-party contest. Wherever possible, results have been gathered from state or local election authorities. Louisiana and California both use nonpartisan blanket (“jungle”) primaries, where all candidates run against each other and (in the event no candidate wins a majority) the highest two vote-getters advance to a run-off. We include the second round run-off election if it features candidates from different parties. Otherwise, we only include the jungle primary. In these cases, the Democratic and Republican vote presented is the combined vote cast for all candidates of each party. If only members of one major party run in the jungle primary, it is considered uncontested. The results of jungle primaries are not included if they yielded a two-party second round. For example, GA-S54 had a jungle primary on December 13. Two candidates from different parties advanced to the run-off on January 10, so we only include the contested second round. The race for GA-S54 featured a Republican and an Independent who we code as a Democrat because of her stated intention to caucus with the Georgia Democratics. The jungle primary results in the congressional CA-34 race is included instead of the runoff which included two Democrats.

Including these jungle primaries and other unusual circumstances does not change the substantive conclusions reported above though the magnitude of underperformance is a little larger if they are excluded but still not statistically significant.

Connecticut employs fusion voting where multiple parties can field the same candidate, so we present the total vote for the Democratic or Republican candidate regardless of ballot party-line.

In the Pennsylvania 197 race, the Democratic candidate missed the filing deadline but went on to run and win as a write-in candidate in a heavily Democratic district. We code this write-in as a Democratic vote.

We use percent of the two-party vote in this analysis. The results are substantively unchanged if we use percentage of total vote or differences between Republican and Democratic vote percentages.

This analysis will be updated as special elections occur through the year and for the regular November state legislative elections in New Jersey and Virginia.