Flip-Flops and High Heels: An Experimental Analysis of Elite Position Change and Gender on Wartime Public Support

Abstract

We address whether politicians’ flip-flopping on support for a war is damaging to their electoral fortunes, and if the gender of the politician has a conditioning effect on this relationship. A series of survey experiments, conducted in 2010 and designed specifically for this project, allows us to examine the causal power of these two cues. Our results challenge the conventional wisdom: respondents do not fault leaders who change their minds about a conflict, and importantly, this effect holds irrespective of the gender of the politician. Instead, individuals react to the policy position the politician currently holds on a war regardless of the politician's consistency and gender.

Keywords:

• conflict
• foreign policy
• public opinion

ACKNOWLEDGMENTS

The authors thank Benjamin Appel, Daniel Corstange, Sarah Fulton, Michael Horowitz, Michael Koch, and Leanne Powner for helpful comments. Data and replication materials can be found at the following website: http://dvn.iq.harvard.edu/dvn/dv/internationalinteractions.

Notes

¹Cited in Koch and Fulton (2011:1).

²Importantly, Tomz and van Houweling (2009) find that candidates who are ambiguous about their policy positions do not suffer negative consequences for changing positions.

³For a review see Karol (2009:25–26). Exceptions include: Allgeier et al. (1979), Croco (2011), McCaul et al. (1995), and Tomz and Van Houweling (2010).

⁴Similarly, Kinder et al. (1980:319) found that “sticks to his word” is a characteristic people associate with a “prototypical president.” Rosenberg et al. (1968) reached a similar conclusion, and noted that respondents rated such traits as “reliable” and “persistent” more favorably than “unreliable” and “wavering.” However, see Sigelman and Sigelman (1986).

⁵See also Poole (2007) and Poole and Rosenthal (1997).

⁶Also see Duerst-Lahti (2006).

⁷Also see Alexander and Andersen (1993), Fox (1997), Kaid et al. (1984), Koch (1999), and Swers (2011).

⁸For example, Duerst-Lahti (2006), Fulton (2012), and Huddy (1994). Kahn (1996:63) found that candidates receiving “stereotypical masculine coverage” fared better than those receiving stereotypical feminine coverage regardless of the actual gender of the candidate.

⁹Given their similarity, we provide a full description of the research design of the first study and elaborate on the changes made for the second study in our discussion of the results.

¹⁰These types of studies have become increasingly common for examining how wartime factors influence mass opinion (for example, Gartner 2008a; Mintz et al. 2006).

¹¹The second study varies information about the senator's party. This is critical since respondents may conflate gender with party (that is, assume women are Democrats and men are Republicans). Making the senator's party explicit helps to isolate partisanship's effect.

¹²We anchored the treatment to an actual event—President Obama's announcement to deploy additional troops to Afghanistan. While the inclusion of Obama's name may trigger a variety of reactions from respondents, we expect these responses to be randomly distributed across treatment groups. Hence, it does not hinder us from drawing accurate inferences about average treatment effects.

¹³In a study, Croco (2011) that employed a similar design but did not address gender, this treatment was sufficient to cue respondents to the flip-flop (or lack thereof). When respondents were asked to elaborate on their evaluations of a senator who changed his position, a sizeable proportion explicitly noted the shift.

¹⁴Control variables are not necessary since we assigned the treatments randomly. This means that any attribute of the respondent that might affect their reaction to the treatment (for example, level of education, interest in politics, etc.) is randomly distributed across treatment groups

¹⁵The fact that the treatment effects for senators who currently agree with the respondent's position are all positive while the parallel effects for those senators who currently disagree are all negative is also noteworthy given the common assumption that foreign policy matters do not play a large role in a citizen's voting calculus. If foreign policy truly did not matter, we would expect small effects across all groups. Yet this is not the case. It could also be true that foreign policy matters a great deal to some respondents and very little to others. This does not affect our ability to draw accurate inferences, however, given that this respondent trait (like any respondent trait) is randomly distributed across treatment groups.

¹⁶As a diagnostic, we interact gender with the other coefficients rather than split the sample. Since this is a non-linear model, we first employ differences to determine if there is a significant change in the likelihood of support of the senator when we move from female respondents to males for each of the different senator types. Only one of the changes was significant: among war supporters, men are more likely than women to think poorly of a female senator who has always been opposed to a war when compared to the baseline (a senator who has always been pro-war). Please note that in a few combinations, the small number of cases led to perfect predictions, making estimations of some interactive effects impossible.

¹⁷ Croco (2011) found the same null effect for consistency employing a national sample on a similar issue. There is not, in other words, any reason to suspect that these findings are limited to student samples.

¹⁸The full set of estimates can be found in Table A1 of the appendix.

¹⁹The one possible deviation from the pattern is the antiwar respondents’ reaction to the female senator who recently joined the antiwar camp; the second bar from the left in Figure 2 appears larger than the first. While this seems to suggest that the respondents are holding the woman to a different standard, there is no difference in a statistical sense between their reaction to her and their reaction to the man with the same voting record.

²⁰Conducting an additional study several months after the first also reduces the likelihood that a single event in the news about Afghanistan might systematically affect the results.

²¹The full set of estimates can be found in Table A2 in the appendix.

²² Levendusky and Horowitz (2012) reach a similar conclusion in their study: respondents had essentially the same reaction to a leader who changed positions, regardless of whether they were from the respondent's party.

²³While Figure 3 implies that gender has no effect on approval, we include a figure in the appendix (Figure A1) that demonstrates this explicitly. The sizes of the effects of moving from a male senator to a female one (while holding other attributes of the senator constant) are uniformly small and the confidence intervals always include zero.

²⁴In a series of t-tests of within each treatment group, we only found a significant difference between genders in two groups: the male and female senators who had been consistent opponents of the war. In both instances, female respondents were more likely to support these senators than male respondents. It is important not to put too much weight in these differences, however, since they are likely an effect of the unbalanced nature of the cells. For the male senator, for instance, there are 10 women, all of whom indicated they would approve of the senator. In the same group, there are only 5 men, 2 of whom said they approve of the senator. This is less than 50%, but this percentage is not meaningful given the small number of observations. A similar pattern emerges when we look at the female senator who was a consistent opponent. There are 15 women in this group, all of whom support the senator. There are 6 men, 4 of whom indicated they support the woman. If male respondents were more likely to refrain from voting for female politicians (and vice versa for female respondents), this is not the pattern we would expect to see. Additionally, we created a diagnostic Same Sex variable in each of the models to investigate whether respondents are more lenient toward same sex candidates. Same Sex is a 1 if the respondent was the same gender as the Senator, and a 0 otherwise. The variable Same Sex was never significant in any of the models; the lowest p-value was .245. We interpret this as further evidence that gender does not have as strong an effect as conventional wisdom would lead us to expect.