# Will "Proven Losers" Cost Democrats Senate Races in '16?

Pop quiz: What do Russ Feingold, Kay Hagan, Michelle Nunn, Mark Begich, Ted Strickland and Joe Sestak have in common?

If you answered that all were highly touted Democratic senatorial or gubernatorial candidates who lost in 2010 or 2014, you would be correct. Feingold, Hagan and Begich lost their Senate seats to Republicans; Nunn and Sestak lost open-seat Senate contests in Georgia and Pennsylvania in 2014 and 2010, respectively; Strickland was unseated as Ohio governor by John Kasich in 2010.

But you would also be correct right if you answered that are all either plausible or presumptive Democratic nominees for 2016 Senate races. Both parties are deep in the candidate recruitment process, and according to AP’s Erica Werner, these six politicians – whom she refers to as “proven losers” – top the Democratic recruiting list in potentially critical Senate contests.

Werner’s (very good) reporting engages with two of the most important questions in the upcoming Democratic Senate primaries: whether these potential candidates’ recent losses signal that they would be weak candidates in 2016, and if renominating them is a sign that their party has a weak “bench” of potential candidates. Both of these questions deserve thorough investigation, but for now I’ll table the latter and focus on the former.

I used data to investigate whether a candidate who lost his or her most recent statewide election bid typically fared worse than other candidates the next time he or she ran. The data showed that, after controlling for a few key election “fundamentals” (like incumbency and presidential approval), being a “proven loser” had no real effect on a Senate candidate’s final share of the vote.

A pretty simple linear regression model produces this result. Put simply, linear regression is a statistical tool that predicts or explains one quantity by using some other quantity or set of quantities. In this case, the linear regression model predicted Senate results from 2002 through 2014 using only a few “fundamentals” – incumbency, partisan lean of the state in the last presidential election, and the incumbent president’s approval rating according to the last Gallup poll before Election Day – and variables indicating which (if either) candidate had lost their last statewide election. Unopposed races, contests with significant third-party vote-getters, and special elections that did not occur on Election Day in 2002, 2004, 2006, 2008, 2010, 2012 or 2014 were excluded.

The goal of this model is not to predict elections perfectly but to discern which variables have an effect on the outcomes after other variables are accounted for. Specifically, the goal is to determine whether a candidate’s “proven loser” status significantly hurt his or her vote share after taking stock of some of the basic election fundamentals. Here’s a trimmed down version of the results*:

The first column of the table lists the variables used to predict the “incumbent party vote share” (where the incumbent party is the party of the president: Republicans were the incumbent party from 2002 to 2008 and Democrats were the incumbent party 2010 through 2014) in Senate contests. The second column lists whether the relationship between that variable and incumbent party vote share is positive or negative. For example, presidential approval has a positive relationship with incumbent party vote share, which means that as the incumbent president’s popularity increases, Senate candidates from his party tend to get a larger share of the vote. The third column lists whether or not the variable was statistically significant – that is, whether this data set confidently says that the variable had an effect on election results.

Note that the partisan lean of the state, incumbency and presidential approval all had significant effects in the direction that one would expect. Incumbent senators tend to get a larger share of the vote than non-incumbent senators. High incumbent presidential ratings lifted candidates from the president’s party, and low ratings dragged them down. And if a state leaned toward the incumbent party (e.g. Alabama leaned towards Republicans in the Bush era; Illinois leaned towards Democrats in the Obama era), candidates from that party tended to gain votes.

But neither of the variables used to indicate whether a candidate was a “proven loser” had a statistically significant effect on final election results. In other words, candidates who lost their last statewide election did not suffer any added penalty in their next election.

These results seem straightforward, but it is important to avoid misinterpreting them. The Democratic Senate candidates in question may still have flaws that both contributed to their previous losses and could follow them into their next election. For instance, one could also argue that Sestak’s difficult relationship with the Pennsylvania Democratic Party contributed to his 2010 loss against now Sen. Pat Toomey and that those strained relationships would harm him again in 2016. Our analysis neither increases nor diminishes the support for that sort of argument.

Instead, our model addresses arguments that run something like “Kay Hagan was a good candidate, but she lost in 2014; what’s to say she won’t blow it again in 2016?” As these primaries progress, party activists and pundits are certain to keep asking this type of question. And our analysis shows that, unlike presidential candidates, Senate candidates who lost once are typically not “losers for life”

*For the most statistically inclined readers – here is a full table of coefficients, standard errors, t-values and p-values for the regression:

The R-squared value is 0.4967. This is just a more detailed delineation of the regression results – readers who are unfamiliar with linear regression do not need to worry about this table.