Affect: The Mediating Mechanism in the Partisan
Divide Over Israel
The large differences between respondents with favorable and unfavorable views of Israel may be rooted in the different feelings of affect that partisans feel toward Israel. If, as we propose, demographic and political sources of attitudes such as party identification explain affect, and affect explains preferences, then affect may be mediating the relationship between sources and attitudes. This is suggested by Page and Bouton (2006) and tested on a range of issues by Gries (2014). We test this proposition, illustrated earlier in Figure 7.1, using our data on the four measures of support for Israel: Sympathies, blame, use of force and foreign aid.
We use the “mediation” package in R (Imai et al., 2010; Imai et al., 2010, 2011; Tingley et al., 2014), which provides a comprehensive framework for examining mediation effects in various statistical models, including, for our purposes, binary logistic regressions. As input for this test, we provide the package with two models for each measure of support. The first model estimates favorability as endogenous to our list of demographic and political predictors. The second model estimates a particular measure of support as endogenous to the same list of predictors, plus the addition of favorability. The models are summarized and explained further in Tables 7.4—7.5 in the Appendix for this chapter. Both models are estimated on the same identical observations.'’
In Figure 7.3 we plot the results of the tests for mediation. The left panel plots the indirect effect of party, through favorability, for each measure of support for Israel. On the right panel, we illustrate the proportion of the effect of party that is mediated through favorability. This is a useful indication of magnitude, pointing to the importance of favorability as a mediator.
The mediation analysis yields an average causal mediation effect (ACME), capturing the indirect effect of party identification. Specifically, we compare the indirect effect of Republicans that is mediated through affect to that of Democrats, while holding all other variables constant. The bars on the left panel summarize the ACME. The black horizontal lines represent 95 percent confidence intervals. The results reveal a significant indirect effect on three of four measures of support. That is, favorability mediates the relationship between party
FIGURE 73 The Mediating Role of Affect
Note: Mediation model estimated using the "mediation" package in R (Imai et al., 2010; Imai et al., 2010, 2011; Tingley et al., 2014). Models estimated while controlling for gender, age, race, religion, party, region and year (as well as any model-specific controls accounting for variation on question wording). The figure includes bootstrapped standard errors for 1,000 samples. Results remain largely unchanged using clustered standard errors by survey (not displayed). Regression estimates are summarized in Tables 7.4 and 7.S in the Appendix for this chapter.
ACME (average causal mediation effect) captures the indirect effect that is mediated through affect. ADE is the average direct effect of party identification on measures of support for Israel. Bars represent the effects. Horizontal lines represent the 95 percent confidence intervals of the effect.
Figure plotted using the ggplot2 package in R (Wickham, 2016).
identification and support for Israel measured in sympathies, Israel’s use of force and providing Israel with aid. The only exception is blame.
The coefficients are small, but given the logistic model, this is unimportant. Rather, the extent to which it is significantly different from zero is important, and perhaps even more so is the magnitude of the mediation, which is captured by the proportion of the effect of party that is mediated through favorability. This is illustrated on the panel on the right. Over half (0.52) of the effect of party on sympathy is mediated through favorability. In other words, so strong is the affect that Republicans feel toward Israel, compared to that of Democrats, that half of the partisan differences in sympathies toward Israel are accounted for through favorability6
The proportion of the mediated effect on other issues is lower but is nonetheless meaningful. On aid and Israel’s use of force, about a quarter (0.26 and 0.27, respectively) of the effect is mediated through favorability.
It is curious that we find no significant effect for blame. Given the significant and large effect of favorability on blame we found earlier in the chapter, we
FIGURE 7.4 Moderating Effect of Year on the Relationship between Affect, Party and Blaming Israel's Adversaries
Note: Mediation model estimated using the “mediation" package in R (Imai et al., 2010; Imai et al., 2010, 2011; Tingley et al., 2014). Models estimated while controlling for gender, age, race, religion, party, region and year (as well as any model-specific controls accounting for variation on question wording). Regression estimates are summarized in Table 7.6 in the Appendix for this chapter.
ACME (average causal mediation effect) captures the indirect effect that is mediated through affect. ADE is the average direct effect of party identification on measures of support for Israel. Bars represent the effects. Horizontal lines represent the 95 percent confidence intervals of the effect (calculated using nonparametric bootstrapping for 1,000 samples).
Figure plotted using the ggplot2 package in R (Wickham, 2016).
attribute this to the first part of the mediation model. In these data, party has no significant effect on favorability, which is unusual. However, data that include both favorability and a measure of blame are rare, and this model relies on only two surveys—one in 1989, when a meaningful difference between Republicans and Democrats was yet to emerge, and a second in 2006 soon after the emergence of partisan differences. Moreover, the majority of observations in this model originate in the earlier survey (1,121 compared to 438 in 2006). When interacting party identification with year,' estimating a moderated mediation model, we find no mediation effect in 1989 but a significant mediation effect in 2006 (see Figure 7.4). In other words, the 1989 data are skewing the results, and favorability does, in fact, mediate the relationship between party and blame in more recent years, when party differences have become more meaningful. According to the 2006 data, 15 percent of the effect of party on blame is mediated through favorability. We list the regression models for this particular test in the Appendix for this chapter as well (Table 7.6).
