Shot by the Messenger: Partisan Cues and Public Opinion Regarding National Security and War

Bacon added, “Other than his opposition to Lieberman's war support, Lamont doesn't have much of a campaign platform.” Following Lieberman’s defeat in the primary, liberal website Dailykos argued, “[Lieberman’s defeat] was also about Lieberman's general desire to do Bush's bidding and to attack fellow Democrats. Which he did full throttle, attacking Lamont for being about just one issue––Iraq, sounding suspiciously like a lot of Republicans in making that charge” (Dailykos.com 2006).

Jentleson and Britton (1998) and Jentleson (1992) find that elite cues––in the form of presidential support or congressional opposition––do influence public support for U.S. conflicts. However, they conclude that the nature of elite rhetoric is endogenous to the principal policy objective, which they argue is a more central causal variable. Moreover, they do not disaggregate public opinion or consider the role of partisan conflict in mediating the effectiveness of elite cues for different groups of citizens.

Individuals’ interpretations of heuristic cues depend in significant measure on their pre-existing belief systems (Hurwitz and Peffley 1987; Herrmann et al. 1997), for which party identification is typically an important (Rahn 1993; Popkin 1994, Lupia and McCubbins 1998; Groeling 2001; Nelson and Garst 2005), albeit incomplete (Holsti 2004), element.

Two related lines of research in social psychology are the influence of “incongruous” (Walster et al. 1966; Koeske and Crano 1968) or “disconfirming” messages (Eagly et al. 1978).

Our predictions vary depending on respondents’ partisan affiliations (Democrats, Republicans, and Independents). Hence, we actually have 24 distinct “cells” of interest in this experiment (8 treatments × 3 partisan subgroups). Given the complexity of this comparison and the limited number of participants, we adopted a randomized comparative experimental structure, rather than incorporating an additional control group that would be unexposed to any treatment. For similar reasons, we interpret our statistical results through a combination of ordinal logit analyses and simulations intended to help the reader more easily interpret and visualize the impact of the treatment conditions across respondent and treatment groups.

Our video treatments use actual news footage re-assembled into new packages designed to maximize realism. Due to a paucity of actual Democratic praise of the president, we were forced to misattribute positive remarks by Sen. Charles Grassley (R-IA) to Sen. Herb Kohl (D-WI), and take other remarks by actual Democrats out of context. We selected Grassley as Kohl because of their relatively low name recognition. For instance, according to one survey, 62% of Americans outside of Iowa had never heard of Grassley (Beaumont 2005). Presumably, only a subset of the remaining 38% would recognize his face or voice. In a separate pilot study, only 11 and 21% of our Democratic and Republican participants, respectively, were willing to rate Grassley on a thermometer scale. The corresponding percentages for Kohl were 21 and 23%, respectively. The remaining rhetoric types were readily available. Still, by using real-world comments, the conclusions we are able to draw from our rhetorical comparisons are somewhat more tentative than would be the case with greater control. This tradeoff did not apply to the static web pages, where statements attributed to members of Congress (MCs) were constant within the praise and criticism categories, and only the identities were changed to reflect the known stances of existing MCs.

It is also possible that differences in the stature or notoriety of individual MCs featured in our treatments could influence their persuasiveness. Our data do not allow a direct test of this conjecture. But they do allow an indirect test. More senior or noteworthy MCs ought, all else equal, to engender more intense, and less neutral, feelings among respondents. After all, such MCs should be more familiar to them. If so, all else equal, we would anticipate finding systematic differences between MCs in the mean distance from the neutral points of their thermometer ratings. Yet the overall average distance from the neutral point across all MCs appearing in our treatments, and across partisan respondents rating them, is less than half of one point (.42 points) on the 0–10 scale, the largest gap across treatments by a given partisan group is about .4 points and the largest gap across partisan respondents’ ratings of the identical treatment is about .31 points. These represent gaps of 3.8, 3.6 and 2.8%. This suggests that our participants had similarly intense feelings toward the MCs featured in each treatment condition, and that these relative intensities were similar across partisan subgroups. This represents at least some suggestive evidence that variations in the stature or notoriety of the MCs in our treatments are not driving our results.

Additionally, recent research has called into at least some question the oft-cited claim that experimental results derived from student subject pools are unrepresentative in important ways. Most notably, Kuhberger (1998) reviewed 136 studies of framing effects and found no significant differences between student and target samples. Our research, though not directly addressing framing, focuses on similar types of cognitive processes. Hence, while it is important to remain cautious in generalizing from a single experimental result based on a single population sample––and especially one drawn from a non-representative subject pool––by the same token, the evidence of a particular systematic bias associated with student population samples, at least in experimental contexts relatively comparable to ours, remains ambiguous.

These figures set responses of “neither approve nor disapprove” to zero. If these responses are set at the mid-point between 0 and 1, overall and national security approval rise to about 7 and 15%, respectively.

We remap “Other” and “None” responses into the Independents category. Including “other” partisans has no significant effect on the ideological orientation of participants in the Independent category.

Most of the controls only modestly affect our results. Yet, given that many are statistically significant––suggesting that, as anticipated, random assignment did not eliminate all bias in our data – we elected to retain them in our final models.

Model 1 excludes four influential outlier observations (.02% of our cases). Including these cases modestly weakens several results and modestly strengthens several others, but does not materially alter the results. We also exclude 16 observations (.09% of our cases) where participants clearly indicated in open-ended questions that they had recognized the treatment manipulations.

For clarity (and brevity) of exposition, we collapse the “strong” and “somewhat” categories in our reported results. Fully disaggregated results are available from the authors.

In a separate analysis (not shown) we also tested for the impact of partisan media outlet credibility. We did so in order to determine whether participants might be inclined to view statements appearing on a network they perceive as ideologically friendly as credible and hence persuasive, while viewing equivalent statements appearing on a “hostile” network as non-credible and hence unpersuasive. Unfortunately, such a model requires a three-way interaction (outlet ideology × viewer ideology × message valence), thereby substantially reducing our statistical leverage. The results from this three-way interaction model are consistent with our predictions, but in some instances at marginal levels of statistical significance. The model focusing on costly credibility allows us to collapse to a simpler two-way interaction, which greatly enhances our statistical leverage as well as simplifying the analysis and discussion of our results. Hence, we focus on the latter model.

Fifty-four responses of “don’t know” or refusals to answer are coded as missing.

This represents about 35% of the sample. Among CNN viewers, 104, 66, and 93 respondents identified themselves as Democrats, Republicans and Independents, respectively. Among FOX viewers, the corresponding numbers are 46, 131 and 66.

PEJ reports that 41 and 39% of FOX and CNN stories, respectively, were neutral, while 15 and 9%, respectively, were categorized as multi-subject and were not coded for tone.

We replicated Model 2 (not shown), first adding a battery of demographic controls (age, education, income, ethnicity, gender) and then a battery of media consumption preference controls (network TV news, Internet news, local TV news, newspapers, CNN). The demographic variables had no discernable effect on respondents’ propensities to watch FOX, or on the three relative correctness indicators. As one might expect, each of the media consumption indicators was highly significant and negatively correlated with propensity to rely on FOX. However, none mediated the effect of the issue correctness measures.

Of course, contemporaneous estimates influence post-hoc retrospective evaluations. The two indicators correlate at .57, indicating that while reasonably strongly related to one another, they are not substitutes.

We include controls for preferring network newscasts or the Internet as sources for national and international political news. Other media outlets (newspapers, magazines, radio) were insignificant and did not affect our results. Hence, they are excluded.

The reported results exclude four influential outlier observations (or .03% of our cases). Including these outliers modestly weakens, but does not materially alter, the reported results.

FOX News chairman Roger Ailes responded by complaining that pressure groups were urging candidates to “only appear on those networks and venues that give them favorable coverage” (Whitcomb 2007).