CAPITULO III: RESULTADOS DE LA INVESTIGACION
3.1. ANALISIS E INTERPRETACION DE LOS DATOS
3.1.2. CONSTRUCCION DEL MODELO INVESTIGATIVO PROPUESTO
To test whether the heightened exposure to conflict experienced by residents of ballot measure states lead to a decline in support for the FMA, I compare the change in the average level of support for the amendment in ballot measure states with the change in the average level of support in non-ballot measures states before and after the referendum process.
measure states) due to the presence of some extreme outliers. Nevertheless, a difference of means test indi- cates the average number of articles in the respective groups are statistically different from one another.
7The exact keywords used for the search are: BODY((gay marriage OR“same sex marriage”) AND (parti-
san OR battle OR argu! OR fight OR conflict OR controvers! OR heated OR feud OR fissure OR intractable OR outrage OR showdown OR standoff OR bitter OR fear OR anger! OR hate! OR assault! OR rage OR raging OR contentious OR condemn OR vehement! OR war))
The analysis is conducted using national survey data from two ABC/Washington Post polls administered by the same survey house in March, 2004 and April, 2005. Both sur- veys asked respondents, “Would you support amending the U.S. Constitution to make it AGAINST THE LAW for homosexual couples to get married anywhere in the U.S., or should each state make its own laws on homosexual marriage?” (ABC News/Washington Post 2004; 2005).8 In 2004, 42% of respondents answered in favor of the amendment and in 2005, 37% did so.9
The timing of these two polls allows the opportunity to assess public opinion before and after the 2004 campaign season. In March of 2004, the referendum process was not yet underway in most states. Utah’s legislature was the first to approve the placement of a marriage referendum on the ballot—the measure passed both houses on March 3rd. Yet, even in this early state, the popular campaign for the measure was in its infancy when the poll went into the field on March 4th. In fact, in 13 of the 15 high conflict states, the issue was decided on or after November 2nd, meaning that the height of the campaigns
surrounding the measures came nearly eight months after the initial March, 2004 poll.10
The 2004 poll, therefore, provides a baseline measure of public opinion regarding the FMA prior to the referendum process.
8The 2004 poll randomly assigned half of the respondents a slightly different question wording. The
alternate wording reads: “Would you support amending the U.S. Constitution to make it ILLEGAL for ho- mosexual couples to get married anywhere in the U.S., or should each state make its own laws on homosexual marriage?” (ABC News/Washington Post 2004, 26). I combine these two question wordings.
9In an analysis of various survey questions on the subject of the FMA, The Pew Research Center finds that
respondents are less likely to favor the FMA when given the option of leaving the matter to the states (Pew Research Center 2004). Levels of support for the FMA reported here may, therefore, be lower than those reported in studies that rely on different question wordings. However, the question wording used does not impede my ability to test the hypotheses I develop as the same question wording was used to assess support for the FMA in 2004 and 2005. I was also sensitive to the possibility that respondents who oppose the FMA are not given a response option that clearly articulates their position on the issue. This could result in high levels of “don’t know” responses among marriage supporters (see Berinsky 2004). I explore this possibility in an analysis provided in Appendix B.2. I find that the question wording does not result in a significantly different percentage of “don’t know” among between marriage supporters and opponents.
10Missouri decided the matter with a popular vote held on August 3rd, 2004, and Louisiana put their
By the time the April, 2005 poll was conducted, residents of ballot measure states had been exposed to the intense debate that accompanies the referendum process. Each of the 2004 ballot measures had been decided by the time the poll went into the field, as had the Kansas marriage amendment that passed on April 5th, 2005. The timing of this second survey allows for an examination of attitudes following the conclusion of the campaign season.
3.5 Analysis and Findings
Figure 3.1 shows the respective percentage of respondents who stated they would sup- port a federal amendment banning same-sex marriage, by year and by residence in a ballot measure state. In 2004, roughly 40% of non-ballot measure state residents supported the FMA and in 2005 approximately 37% of respondents continued to support it. In ballot mea- sure states, however, the change in support for the FMA was much larger. While roughly 48% of residents in these states supported the FMA in 2004, just over 36% stated they supported the FMA in April, 2005—a difference of approximately 12 percentage points.
With a quasi-experimental design like this one, assignment to the treatment group (i.e. referendum states) and control group (i.e. non-referendum states) is nonrandom. States considering amendments banning same-sex unions are likely to have a higher proportion of residents who oppose gay marriage, a higher proportion of conservative residents, and a lower proportion of liberal residents as compared with other states. The resulting imbalance between the treatment and control groups has the potential to cause estimation error and bias in the estimates drawn from these data.
To address the problem of imbalance, I use coarsened exact matching (CEM) to prepro- cess the data prior to running a logit model estimating the likelihood of support for the FMA (Iacus, King and Porro 2011a; Iacus, King and Porro 2011b). Observations in the treatment and control groups are matched on self-reported ideology (measured with two variables, one identifying liberals and one identifying conservatives), attitude toward gay marriage,
2004 2004 2005 2005 0 10 20 30 40 50 60
No Ballot Measure Ballot Measure
Per cen t o f R esi d en ts in Su p p o rt o f F M A
Figure 3.1: Levels of Support for the Federal Marriage Amendment Between 2004 and 2005 in Ballot Measure and Non-Ballot Measure States
and year. Variables measuring attitudes toward gay marriage are constructed from ques- tions that gauge support for the legalization of same-sex marriage and civil unions.11 Two
such variables are created, one that indicates opposition to marriage and civil unions, and one that indicates support for civil unions but not marriage.12 This process “prune[s] ob- servations from the data so that ... the empirical distributions of the covariates (X) in the groups are more similar” (Iacus, King and Porro 2011a, 1). Preprocessing the data with matching also has the advantage of reducing model depedence (Ho, Imai, King and Stuart 2007).
11The question wordings are provided in Appendix B.3.
12See Appendix B.2. One might anticipate that attitudes toward same-sex marriage would perfectly predict
support for the FMA. In actuality, however, of the 868 individuals who responded that same-sex marriage should not be legal, 257 (roughly 30%) responded that the states should make their own marriage laws.
Prior to matching, the data had a multivariate L1 distance (a measure of global imbal-
ance) of approximately .13. A multivariate L1distance of zero indicates perfectly balanced
data. After matching, the rounded multivariate L1 distance is zero.13 With data that are
exactly balanced (as these now are), control variables are not needed to establish treatment effects, as the control variables are no longer related to the treatment group (King et al. 2011a).
To determine whether the change in support for the FMA among ballot measure state residents differs significantly from the change among non-ballot measure states, a logit model is used to estimate the likelihood of support for the FMA. The data are weighted with the CEM weights generated by the CEM algorithm and the standard errors are clus- tered by state.14 The dependent variable is a dummy coded as one for respondents who support the FMA and zero otherwise.15 The key independent variables of interest are: a year dummy (coded as zero for 2004 and one for 2005), a dummy representing residence in a ballot measure state, and an interaction term that multiplies the two. The coefficient on the interaction term gives the difference-in-difference estimate for the treatment effect of residence in a ballot measure state.
Table 3.3 displays the results of the analysis. The coefficient for the interaction term “Ballot Measure State X Year” is negative and statistically significant, indicating that sup- port for the FMA decreased among ballot measure state residents between the spring of 2004 and the spring of 2005.
To assess the magnitude of the treatment effect in substantive terms, Figure 3.2 dis- plays the predicted probability of support for the FMA by year and by residence in a ballot
13The precise multivariate L
1distance is 1.287e-15.
14Hierarchical linear models (provided in Appendix B.4) returned nearly identical results and support
the same conclusions as do the logit models presented. However, HLM models do not support importance weights in STATA, and therefore, can not utilize matched data. For this reason, I find that logit models with clustered standard errors are preferable.
15The zero category of the dependent variable includes both individuals who prefer to leave the matter to
measure state. The first set of predicted probabilities (plotted in black) shows that the prob- ability of support among ballot measure state residents dropped from nearly 48% in 2004 to approximately 36% percent in 2005—a difference of 12 percentage points. The confi- dence intervals for these two sets of predicted probabilities do not overlap, indicating that the change in support for the FMA is statistically significant. The probability of support among residents of non-ballot measure states (plotted in gray) decreased by just two per- centage point during this time and the change is not statistically significant—as indicated by the overlapping confidence intervals. These results provide strong evidence that the de- bate over gay marriage that accompanied the referendum process lessened support for the FMA in ballot measure states, while support in other states remained stable.
Table 3.3: Model A—Likelihood of Support for the FMA by Year and Residence in a Bal- lot Measure State
Variable Model A
Ballot Measure State Dummy 0.10 (0.15)
Year -0.08
(0.09)
Ballot Measure State X Year -0.41*
(0.16)
Constant -0.17
(0.09)
N 2165
AIC 3010.69
Standard errors (in parentheses) are clustered by state.
* Indicates statistical significance at the level of 95% confidence.