The conditional influence of an endorsement

Recently, American politics scholarship has undergone a paradigm shift in conceptualizing political parties. The once heavily candidate-centered focus on individuals taking the initiative to seek office under the party label, with the party acting in service to these political entrepreneurs (Aldrich Reference Aldrich1995), has fallen out of favor. To be sure, there remains credibility in viewing the major parties from this vantage, but it is also the case that as Democrats and Republicans polarized and became more driven by ideological differences (Abramowitz Reference Abramowitz2018) and affective dissension (Mason Reference Mason2018), a framework built around the claim that parties are coalitions of various groups of policy demanders has gained theoretical and empirical credence (Bawn et al. Reference Bawn, Cohen, Karol, Masket, Noel and Zaller2012; Cohen et al. Reference Cohen, Karol, Noel and Zaller2008). According to this perspective, the growing influence of the parties is evident in their ability to direct co-partisans in the electorate to support their choice in primary/nomination contests, whether for president or lesser offices. One tangible measure of party influence is the use of endorsements to attract electoral support in contested primaries. Whether it be a presidential nomination (Cohen et al. Reference Cohen, Karol, Noel and Zaller2008), a House (Dominguez Reference Dominguez2011), or Senate primary (Hassell Reference Hassell2016), party-generated endorsements (and/or financial contributions) are correlated with greater electoral backing in these contests.Footnote ¹

Nevertheless, the “party decides” model of American political parties runs into complications if in fact the party is unable to coalesce around a single candidate and, even worse, if the candidate most elites in the party network oppose actually wins the nomination. This is of course the scenario that played out in 2016 when the political outsider Donald Trump won the GOP presidential nomination. We contend that Trump was a “movement” candidate because he captured the support of most Republican voters despite the efforts and signaling of numerous GOP elites who tried to make a case against him.Footnote ² As MacWilliams (Reference MacWilliams2016) sees it, because the Republican establishment failed to settle on an alternative to Trump, his unique appeal, greatly rooted in authoritarianism, propelled him to the 2016 Republican presidential nomination. Hence, as Hassell (Reference Hassell2016, 76) points out, “Just because party networks want to influence primaries does not mean that they have the resources or power to do so.” They did not in 2016, and in winning the White House, Trump took control of the Republican Party.

Manento and Testa (Reference Manento and Testa2022, 1654) pose an interesting question: “When a party outsider captures the party, who is then considered an outsider?” By the time former president Trump issued a slew of candidate endorsements in the 2022 midterm cycle, by dint of his enduring popularity among Republican voters and his remaining engaged in the political arena, Trump persisted as the de facto head of the GOP (Jacobson Reference Jacobson2021). Thus, in a remarkable bottom-up political twist, by achieving historically high approval among rank-and-file GOP voters, the Republican Party network bowed to Trump. With his passionate base of supporters and far-reaching influence within the GOP, Trump deployed Republican primary endorsements to continue leveraging current and future political concessions within the party he leads.Footnote ³ Given this reality, we should expect Trump’s endorsement to be especially impactful in GOP primary contests, and there is evidence it was (Manento and Testa Reference Manento and Testa2022).

However, context is critical in thinking about the significance of a Trump endorsement. Georgia was a setting where a Trump endorsement should be highly variable because of the ex-president’s controversial involvement with Republican officials and election administration, after he lost the state to Joe Biden (Hood and McKee Reference Hood and McKee2022a; Hood and McKee Reference Hood and McKee2023a). We now turn to a discussion of the likely conditional effects of a Trump endorsement based on what we know about political behavior.

As stated by Manento and Testa (Reference Manento and Testa2022, 1634), “In an age of popular primary elections, for party organizations to see their preferred candidates nominated, the voters must act in accordance with the wishes of party elites. Yet it is not clear that voters always follow the cues of the party.” The American voter is notoriously disinterested in politics, possessing a myopic view of current events, which places hard limits on their depth of decision-making and hence a frequently faulty capacity to properly assign electoral punishment and reward (Achen and Bartels Reference Achen and Bartels2016). To be sure, there is a small stratum of politically sophisticated ideologues and issue publics (Converse Reference Converse and Apter1964); the former care to know about politics because it interests them (Luskin Reference Luskin1990), and the latter easily acquire information because it directly impacts their livelihoods. But for most of the mass public, cue-taking and shortcuts are greatly valued devices because it is an efficient means to engage a subject most could easily care less about (Lupia Reference Lupia1994; Popkin Reference Popkin1991).

There is an interesting irony to the context of primary elections. They attract the most engaged voters, at least by evidence of their willingness to participate, even if this smaller contingent is typically representative of the general electorate sharing the same party affiliation (Geer Reference Geer1988; Sides et al. Reference Sides, Tausanovitch, Vavreck and Warshaw2020). But the choice calculus among these more participatory voters is considerably more complicated at the primary stage when there exist multiple candidates all affiliated with the same party (Bartels Reference Bartels1988). With the party label constant and the presence of several candidates, an endorsement from a high-profile actor can be a helpful and efficient tool for primary voters to draw upon in forming their decision (Grossman and Helpman Reference Grossman and Helpman1999; Pease and Brewer Reference Pease and Brewer2008; Garthwaite and Moore Reference Garthwaite and Moore2013). Enter Donald Trump, a businessman, celebrity, and political figure whose presidential popularity among Republican voters achieved all-time highs while also registering the most polarized partisan disparities in approval ever recorded by the Gallup Poll (Jacobson Reference Jacobson2021; McKee, Evans, and Clark Reference McKee, Evans and Clark2022). With universal name recognition and an enduring hold over the Grand Old Party because of his continued involvement in politics after losing reelection in 2020, a Trump endorsement should have been influential in Republican primary contests.

Nevertheless, there are natural limits to the influence of any endorsement, even one made by Republican voters’ most approved of president. Foremost, voters weigh the endorsement against prior information they have on primary candidates. This means that a Trump endorsement should exhibit a highly variable effect on vote choice. Hence, we expect a Trump endorsement registered its greatest effect in the lowest-profile primary contests. In the absence of any competing information on primary candidates, the positive association with a Trump endorsement should have moved most voters to favor Trump’s endorsee, which also necessarily reduced the share of otherwise undecided voters. Thus, in a Republican primary the Trump endorsement essentially served the dual purpose of persuading and informing. However, a Trump endorsement should have exhibited negligible effects in the most high-profile primary contests because voters possess more information that makes it easier to shape their candidate preference regardless of Trump’s choice. For instance, a well-known incumbent cultivated a favorable profile with voters, and their impression of this politician is likely developed to the point that any endorsement, even a Trump endorsement, exhibited at best a minimal to modest influence on vote choice.

Table 1 shows summary information for five statewide offices on the May 2022 Georgia Republican primary ballot in which Trump gave his endorsement. First, the number of candidates running in these contests ranges from three (insurance commissioner) to six (US Senate). Second, the salience of these contests was remarkably variable; with few voters possessing information on insurance commissioner candidates, even with the presence of an incumbent officeholder appointed by Governor Kemp. In contrast, well-known Governor Kemp drew four challengers for renomination, including Trump-endorsed David Perdue, who most recently lost reelection to the US Senate in an early January 2021 runoff against Democrat Jon Ossoff.Footnote ⁴

Table 1. Summary information for Trump-endorsed Georgia GOP statewide primary contests

Note. Trump also made an endorsement for attorney general, but this came too late to have included the race in our survey experiment. Among the five contests in the table, the most votes were cast for governor (1,203,835). Roll-off, the percentage less votes cast in the remaining contests, was as follows: 2% for US Senate; 3% for secretary of state; 7% for lieutenant governor; and 8% for insurance commissioner. Data are from the Georgia Secretary of State’s website.

Table 1 also displays the vote share garnered by each candidate in each primary race, denotes the incumbent in bold if there is an incumbent running, indicates the Trump endorsee in the contest, whether Trump’s endorsed candidate won, and the relative profile of the race. Notably, the significance of incumbency appears present (Ansolabehere et al. Reference Ansolabehere, Hirano, Snyder and Ueda2006) at least with respect to the fact that in the three races with an incumbent, all three won renomination. Trump did not endorse these incumbents and his endorsed candidates finished second in each of these races – an indication that the endorsement likely had some influence but not enough to change the outcome. By comparison, in the two open contests (US Senate and lieutenant governor), Trump backed the winner. Although the incumbency advantage is trending toward an all-time low (Jacobson Reference Jacobson2015) because of the nationalization of American voting behavior (Bartels Reference Bartels2000; Hopkins Reference Hopkins2018), it probably retained some potency in Georgia because the ballot indicated the incumbent in the race, which served as a useful cue particularly in a primary race that featured candidates of the same party (Schaffner, Streb, and Wright Reference Schaffner, Streb and Wright2001).Footnote ⁵

The last column of Table 1 holds particular importance when thinking about the significance of an endorsement. Again, we hypothesize that a Trump endorsement mattered more in low-profile contests. Given Trump’s high popularity among an electorate disproportionately comprised of GOP identifiers, if they have scant information about the candidates in a certain contest, then Trump’s endorsement was a potent positive cue to guide the vote decision. Further, with minimal information in a low-profile race featuring multiple contenders, the Trump endorsement not only informed the vote decision but, in the process, greatly reduced the number of voters who otherwise would register a Don’t Know response in a more obscure contest.

In determining the level of profile in the five contests in Table 1, a reliable metric is to consider voter roll-off. By this account, the gubernatorial race was the highest profile as it received the greatest number of votes (see the note under Table 1). The US Senate contest was not far behind with 2% less votes cast (so it is also labeled high profile). Because of the attention cast on the secretary of state race, with Trump vehemently opposed to the incumbent Brad Raffensperger, who refused to accommodate the president’s “request” for finding enough votes to overturn Biden’s Georgia victory, we rated this contest as one of a moderate profile (3% roll-off vis-à-vis the gubernatorial race). Lastly, the two lowest-profile races are lieutenant governor and insurance commissioner, with voter roll-off of 7% and 8%, respectively.

Finally, in theory, a Trump endorsement should weigh heaviest in lowest-profile races. However, what complicates the possible influence versus actual influence of Trump’s endorsement is whether primary voters are indeed informed of it. For example, Trump elevated the profile of the secretary of state race because of his quarrel with Raffensperger, and his endorsement of challenger and Congressman Jody Hice appears to have greatly reduced the chances of the incumbent winning.Footnote ⁶ By comparison, in the highest-profile race for governor, incumbent Brian Kemp was very well-known and popular for his record of accomplishments. In this context, as we will demonstrate, the Trump endorsement of challenger and former Senator David Perdue had basically no electoral payout. And in the race where a Trump endorsement should matter most, for insurance commissioner, because of the obscurity of the candidates (Maisel Reference Maisel1982), the incumbent sailed to victory (71% of the vote). But, as we will show, the effect of a Trump endorsement in the experimental setting was greatest in this contest. The problem is that there is limited external validity because such a low-profile contest as insurance commissioner meant that few Republican primary voters were aware of the Trump endorsement.

Data and methods

This preregistered experiment was conducted through the University of Georgia’s School of Public and International Affairs (SPIA) Survey Research Center from March 20 to April 8, 2022 (Hood and McKee Reference Hood and McKee2023b). A total of 736 likely Republican primary votersFootnote ⁷ in Georgia were interviewed over the phone using live callers.Footnote ⁸ Respondents were randomly assigned to either the control group or the treatment group at the beginning of each interview. There were 391 respondents in the treatment group (53%) and 345 respondents in the control group (47%).Footnote ⁹ The control group was asked who they planned on voting for in the following five statewide contests: governor, US Senate, lieutenant governor, secretary of state, and insurance commissioner. Respondents in the treatment group were told which candidate former president Trump had endorsed prior to being asked their vote preferences. All respondents were then asked a set of additional political and demographic questions.

Our analysis proceeds in two parts. First, we display the results for each of the five races analyzed (see Table 2). Here we examine the difference in the estimated vote share for the Trump endorsee across the treatment and control groups. We also display results for undecideds and the remaining candidates. Second, we construct a series of multinomial logit models, by office. Because of the nature of this randomized survey experiment, it is not necessary to include controls.Footnote ¹⁰ For the multivariate analysis (see Table 3), we construct models by office, where the dependent variable is coded 2 to indicate support for the Trump endorsee, 1 for all other candidates in the race, and 0 for undecideds (those answering Don’t Know when asked who they planned to support).Footnote ¹¹ Given the unordered nature of the dependent variable, models were estimated using multinomial logit with the undecided category serving as the baseline comparison. Following estimation, we calculated simulated probabilities (Table 4) and associated confidence intervals for each of the three categories of the dependent variable (Table A7 in the Supplementary Material), across the treatment and control groups. The primary variable of interest for our analysis is the respondent’s assignment to the control or treatment group where the indicator Treatment is coded 1 for those in the treatment group and 0 for respondents in the control group.

Table 2. Support for Georgia GOP primary candidates by Trump endorsement

Note. The Difference column is calculated as Treatment minus Control.

^a Probability difference between treatment and control groups is significant at the 0.05 level.

Table 3. Predicting support for Trump endorsees in the 2022 Georgia GOP primary

Note. Entries are multinomial logit coefficients with standard errors in parentheses. DV: Trump Endorsee = 2; All other candidates = 1; and Don’t Know = 0 (Base Category).

* p < 0.05.

^** p < 0.01.

^*** p < 0.001.

Table 4. Support for Georgia GOP primary candidates by Trump endorsement

^a Probability difference between treatment and control groups (Treatment minus Control) is significant at the 0.05 level.

Findings

As we have discussed, the information environment of a primary election can be challenging. Once the voter gets past top-of-the-ballot contests like governor and US Senate, the low profile of the other offices, and hence dearth of knowledge on candidates running in them, makes a salient cue like Trump’s endorsement a powerful factor in guiding vote choice. We begin by discussing the findings from Table 2.

Starting with the gubernatorial contest, we see that Trump’s endorsement barely moved the needle in favor of former senator David Perdue. His vote share increased two points over the control group. By comparison, in the US Senate primary, Herschel Walker’s vote share among the treatment group increased 11.6 points. Nevertheless, it is also clear that Walker could have easily won without Trump’s endorsement because he garnered just over 64% of the vote among the control group. Most of the increase in Walker’s vote came from a 9-point drop in the Undecided category. In contrast to many national media reports, it is clear that Trump’s endorsement of Walker was not the overriding factor in his winning the GOP nomination for US Senate.

The remaining three down-ticket races reveal a much more pronounced effect for Trump’s endorsement, with the difference between the control and treatment groups ranging from 29 to almost 44 points. In the lieutenant governor’s race, the Trump endorsee Burt Jones’s vote share increased from 30% to 59%, while the vote share for the other candidates dropped a modest 4.8 points. Most of the movement comes from the Undecided category, which dropped from 54% for the control group to 30% for the treatment group.

In the secretary of state contest, Trump endorsee Jody Hice’s vote share doubled from 30% to 60%. The vote share for other candidates in the race, which included incumbent Brad Raffensperger, dropped a more modest 9.5 points, while the share of undecideds fell just over 20 points, from 39% to 18.5%. Finally, in the race for insurance commissioner, the Trump endorsee Patrick Witt’s estimated vote share skyrocketed from 8% to 52% – a difference of nearly 44 points. Most of this movement was related to the concomitant decline in the percentage of undecideds, going from 71% to under 37%.

Table 3 shows the results of our multinomial logit models. Again, the baseline category for comparison is the group of respondents who were undecided when asked about their vote preference. First, we will examine the results comparing those who intended to vote for a candidate who was not the Trump endorsee. The treatment variable, not surprisingly, does not exert a statistically significant effect on differentiating between these respondents. In contrast, comparing those respondents supporting Trump-endorsed candidates to the base category, we see that for all but the governor’s race, the treatment variable is positive and significant. Respondents exposed to Trump’s endorsement were more likely to express support for these candidates.

Multinomial logit coefficients are difficult to interpret directly. As such, we converted the model coefficients into a set of predicted probabilities using the observed-value approach as suggested by Hanmer and Kalkan (Reference Hanmer and Kalkan2013). The predicted probabilities for the five statewide contests are displayed in Table 4 and Figure 1.Footnote ¹² Not surprisingly, the results of the multinomial logit models comparing support for the Trump-endorsed candidate between the treatment and control groups are very similar to the descriptive findings in Table 2.

Figure 1. Support for Georgia GOP primary candidates by Trump endorsement.

In the governor’s contest, the difference between the treatment and control groups in support of the Trump endorsee, at 2.0, is negligible and not statistically significant. At 8.9, the difference for the US Senate contest is statistically significant, but certainly more modest as compared to the three down-ticket races. In the three contests for lieutenant governor, secretary of state, and insurance commissioner, there is evidence of a decline in the likelihood of voting for the candidates not endorsed by Trump. However, the greatest impact of being exposed to the Trump endorsement (a respondent in the treatment group) is reflected in the large increase in the probability of preferring the Trump endorsee and a decline in the likelihood of being undecided (choosing the Don’t Know option).

The difference in the likelihood of preferring the Trump endorsee, depending on whether the respondent is in the treatment group versus the control group, for lieutenant governor, secretary of state, and insurance commissioner, respectively is 0.29, 0.30, and 0.43. Clearly, these are substantial and statistically significant disparities that emerge from whether one is provided information on the contender Trump endorsed. Likewise, the decline in the probability of being undecided based on whether one is in the control group vis-à-vis the treatment group, for lieutenant governor, secretary of state, and insurance commissioner, respectively is 0.24, 0.20, and 0.34. In line with our contention that the secretary of state contest was higher profile than the other two races, the difference in the undecided category is the lowest. Thus, residing in the treatment group obviously results in a substantial reduction in the likelihood that a respondent selects the Don’t Know option in down-ballot contests for lieutenant governor, secretary of state, and insurance commissioner.

Figure 1 provides a visual display of the set of predicted probabilities generated from our multinomial logit models in Table 4. For each GOP primary contest, we plot the predicted probabilities for the treatment and control groups for Trump’s endorsed candidate, other candidates, and those who indicated they were undecided. Each horizontal probability bar also contains its respective 95% confidence interval. This presentation of the data highlights the effect of being exposed to the treatment condition, with palpable shifts in support for Trump’s endorsed candidate and attendant declines in undecided voters, especially in the three lowest-profile races for lieutenant governor, secretary of state, and insurance commissioner.