What I Like About You: Legislator Personality and Legislator Approval

Primary elections (Brady et al. 2007) mitigate these effects in the general election, largely because of ideological convergence concerns among the primary electorate.

Enelow and Hinich (1982) further discretized valence attributes, breaking them down into valence attributes—immutable attributes on which all voters hold the same evaluation—and ascriptive attributes—immutable attributes on which the evaluations of voters differ—with one of the primary distinctions between the types of attributes being that candidates can choose position attributes, while valence and ascriptive attributes cannot be selected.

The underlying basic preferences captured by personality measures such as the Big Five should influence the emphases legislators put on their careers—at home or in Washington—and perhaps the relative costs of those leadership styles. However, other influences, such as those mentioned by Fenno (1978), are mitigating factors as politicians choose how to allocate their resources during their career. While home style may be highly influenced by personality, we would not go so far as to venture that home style is personality.

Moreover, it is not just voters who care about the competence, integrity, effort, and effectiveness of officeholders. When choosing those appointees ultimately responsible for policy implementation, executives have long weighed competence—and other nonpolicy factors—as important considerations necessary to effective policymaking (Aberbach and Rockman 2000; Gailmard and Patty 2007; Hollibaugh et al. 2014; Hollibaugh 2015a, b; Huber and McCarty 2004; Lewis 2008, 2009; Moe 1985; Nathan 1983).

Also see McCurley and Mondak (1995).

There may be links between voter personality and legislator approval beyond personality valence or congruence. For example, voters with certain personality types may differ in the policy preferences they desire from legislators, following the argument of Caprara and Zimbardo (2004). Additionally, certain Big Five personality types may exhibit different preferences over legislator valence characteristics such as character or leadership ability. These questions are worthy of further examination, but are beyond the representation focus of this project.

While the TIPI is shorter than standard instruments, it is well-suited to tasks like the CCES, where time is limited, and results from the TIPI tend to be highly correlated with the results one would get from longer and larger batteries of questions (Ehrhart et al. 2009; Gosling et al. 2003). The question wording is in Online Appendix I.

One concern that arises here is that we are relying on respondents to sincerely place the personalities of their elected officials. Specifically, one might be concerned that there might be a halo effect, as suggested by Caprara et al. (2007), whereby respondents assign “good” personality types to legislators who they already like. This can induce a form of simultaneity bias. We show in Online Appendix L that this concern is not substantial. If present, such a bias is shown in the Online Appendix to be always downward in nature.

Unfortunately, because ELUCIDATION scores do not exist past 2014, we do not include the 2016 CCES in our validation procedure. Moreover, due to the way ELUCIDATION scores are calculated, we do not have any scores for those whose first—or only—session of Congress was the 113th.

Both ideology and personality questions utilize seven-point scales, thus ensuring some degree of scale comparability for our ideological and personality measures.

There may be some concerns that respondents may differ in their abilities to rate legislator personality and may exhibit different relationships between legislator personality ratings and approval on the basis of respondent political information. In Online Appendices G and H, we separately examine high and low information voters and find little difference between the two groups with respect to the relationship between legislator personality and approval.

Our measures are subjective and thus based entirely on the perceptions of respondents about their own personality and the personalities of their legislators. While it would be interesting to examine the relationship between personality congruence and valence between voters and elected officials as measured by experts or longer, more reliable inventories, our objective is to test for possible associations between job approval and perceived personality congruence/valence in the minds of voters, and for this task our measures are very well suited.

Despite the previous validity checks, there might be some residual concern that endogeneity plagues our analyses. However, the structure of the 2014 CCES somewhat mitigates this concern. For the Representatives and half of the Senators, the TIPI questions were asked prior to the approval questions, and both were asked during the pre-election wave. However, for the remaining Senators (denoted here as “Senators 2”), the TIPI questions were asked during the post-election wave, and the approval questions were asked during the pre-election wave. While this setup (unfortunately) forces us to use post-election questions as explanatory variables where a pre-election question is the dependent variable, separate regressions where the two Senators are estimated separately (for the 2014 panel) provide substantively identical results, though the results for are somewhat weaker for the Senator 2 analysis (see Online Appendix D). These results not only suggest that there are no serious problems with using a post-election variable as an independent variable in the present case, but also that the aformentioned endogeneity issue is of little practical concern.

It is possible that causality may flow in the opposite direction and voters may state that politicians of whom they already approve display “positive” personality types. In this case, it would only uncover valence properties of incumbent personality through another pathway, as these voters are implicitly stating their preference for elected officials with these personality trait profiles. This is more likely when voters have low amounts of information about candidates, but in this project we are working with evaluations of incumbent legislators, making this unlikely. That said, we have also conducted analyses using voters’ evaulations of President Obama (using the 2016 sample, since voters were not asked to rate the personality of President Obama in the 2014 sample). While the relationships between voters’ evaluations of President Obama’s personality traits and their approval evaluations are somewhat noiser than the legislator models (perhaps due to the smaller sample or less malleable opinions about a President in his eighth year), the overall findings for the Obama-specific models are substantively similar to those found for the Congressional models. These results are in Online Appendix C.

Also see Hollibaugh et al. (2012).

However, since party likely plays a strong role independent of ideology, we replicate our main results focusing only on those legislator-respondent dyads where the legislators and respondents are of the same party, as well as a separate subset where the legislators and respondents are not of the same party. These results, which are in Online Appendices E (copartisans) and F (outpartisans), are substantively similar to those presented in the main body of the paper.

While theoretically possible, we do not perform Ramey-scaling on the personality variables. This is for several reasons. First, both the Aldrich and McKelvey (1977) method, as well as Ramey’s (2016) modification thereof, require that the analyst choose an individual stimulus a priori to anchor the recovered space. In the context of ideology, this is relatively straightforward, since many reliable measures of ideology exist (e.g., Clinton et al. 2004; Poole and Rosenthal 1997) and only one dimension (the underlying liberal-conservative dimension) needs to be estimated. In the context of personality scores, reliable outside measures of all personality traits are scarce, and even the ELUCIDATION scores might be somewhat problematic for this purpose, since while they correlate with respondents’ perceptions of personality, the relationship are quite noisy; further recall that there is no relationship between the ELUCIDATION score for Conscientiousness and respondents’ perceptions of how their legislators rate on this trait, which would make it particularly hard to anchor it to any degree of confidence. Moreover, while the Aldrich and McKelvey (1977) method and Ramey’s (2016) modification thereof account for within-respondent correlations across stimuli, they do so in a circumstance where only one latent dimension is being recovered. There is undoubtedly within-subject correlation across the five latent personality dimensions, and properly accounting for that within the broader Aldrich and McKelvey (1977) framework would require the development of an entirely new method, which is beyond the scope of this article. Finally, there are substantive concerns as well. The extant literature on the stability of personality traits over time suggests that the expression of persistent personality trait dispositions is conditioned by stochastic-contextual processes as well the social environments and roles to which individuals are exposed (Fraley and Roberts 2005). For example, an introvert may be quite talkative in the context of a family dinner while exhibiting a reserved demeanor among strangers. The assignment of context by nature is inherently noisy and not necessarily random. As a result, several researchers have referred to personality measures as exhibiting ‘patterns of stability’ or ‘patterns of continuity’ and have accepted a significant degree of variation in these measures (McCrae and Costa 1994; Fraley and Roberts 2005). This inherent variation in any measure of personality makes it particularly difficult to defend the choice of any extreme anchor, and by extension, the use of the Aldrich and McKelvey (1977) method for personality traits.

We omit the respondent and office-year random effects from office-specific models and Obama models due to the lack of variation in these cases. In these cases, the office-year random effects simply reduce to year-specific random effects, and the lack of variation—each intercept can only take one of two values in these cases—leads to convergence issues. Thus, in order to avoid these problems, we simply omit these random effects.

Note that the Representative-specific model implicitly accounts for cycle effects, since all Representatives are up for reelection every two years.

Both of these tests are described in greater detail in Online Appendices J (Clarke Test) and K (Vuong Test).

In Monte Carlo experiments (Clarke 2007), the Vuong (1989) test was shown to be relatively conservative and to choose the “correct” model less often than Clarke’s (2007) distribution-free test. However, the Clarke (2007) test chooses the “wrong” model more often. Indeed, the Vuong (1989) test’s relative conservatism prevents it from declaring either model to be the “correct” model in a large proportion of cases; when it does choose a model, however, it tends to be correct. The Clarke (2007) test is more powerful and more likely to declare one model fits better than the other, but it also tends to choose the wrong model more often. However, the same Monte Carlo results suggest neither test chooses the wrong model often in absolute terms.

We focus on the congruence variables here because they are more likely to be affected by measurement error, and the encompassing models will allow us to determine whether any useful signal is buried under the noise—which would likely be the case if congruence were at all predictive of legislator approval in a substantive sense—or whether congruence is of minimal to no consequence.

The encompassing models themselves are in Online Appendix B.

However, in some sense the substantive interpretation of the valence-based variables are of secondary importance, relative to the finding that valence seems to outperform congruence on nearly every metric.

For Fig. 4, the estimated coefficients from the “Ideological Divergence, Personality-Respondent Ideology Interaction, and Political Information” models (Models A-4, A-9, and A-14) are used (see Online Appendix A), as these models are the valence-based models with the lowest BIC among those that include Ideological Divergence as a control variable.

The results presented are for the “average” respondent, in that the random effects terms are set to zero.

However, the capacity for altruism can also work in favor of politically divisive tactics. For example, Ramey et al. (2017) find that, in many cases, Agreeableness manifests as a willingness to put one’s own goals and preferences aside for the good of one’s party.

In Fig. 5, all covariates—except for the respondent ideology (and its absolute value) and the personality trait allowed to vary—are set to zero. Respondent Ideology was set to one standard deviation below its mean for “liberal” respondents, at its mean for “moderate” respondents, and to one standard deviation above its mean for “conservative” respondents. This forces the model to consider the situation where both Senators are identical on every dimension, with the exception of the personality trait under analysis. As before, the results presented are for the “average” respondent, in that the random effects terms are set to zero. Finally, for all plots, the estimated coefficients from the “Ideological Divergence, Personality-Respondent Ideology Interaction, and Political Information” model (Model A-34) are used (see Online Appendix A), as this model is the relative valence-based model with the lowest BIC among those that include Ideological Divergence as a control variable.

Separately valence models were estimated for high and low information voters in Online Appendices G and H. Among ideologically moderate voters in the models pooling all officeholders, high information voters exhibited strong valence preferences for Openness, Conscientiousness, and Agreeableness in legislators, while low information voters exhibited somewhat weaker preferences for these three traits, but also a valence preference for Emotional Stability.

See http://​www.​gallup.​com/​poll/​1597/​confidence-institutions.​aspx.