Gay Rights in Congress: Public Opinion and (Mis)Representation

Transcription

1 Gay Rights in Congress: Public Opinion and (Mis)Representation Katherine L. Krimmel Jeffrey R. Lax Justin H. Phillips Department of Political Science Columbia University September 8, 2012 Abstract Public majorities have supported several gay rights policies for some time, yet Congress has responded slowly if at all. We address this puzzle through dyadic analysis of the opinionvote relationship on 23 roll call votes between 1993 and 2010, matching members of Congress to policy-specific opinion in their state or district. We extend the MRP opinion estimation technique so that it can be more often used for district-level analysis. We find that policy-specific opinion is a very strong determinant of legislator voting, but find large gaps in responsiveness and biases in policymaking. While opinion strongly influences white male Democrats, black lawmakers and white female Democratic lawmakers generally cast pro-gay votes, while Republicans consistently oppose gay rights, regardless of constituent preferences. We also unpack polarization over time, showing the Democrats moving into and the Republicans out of sync with their constituents. This yields a broader, deeper picture of the opinion-vote relationship.

2 1 Introduction Over the past two decades, public support for gay and lesbian rights has increased dramatically. Indeed, many pro-gay policies now see majority support, even in conservative areas of the country. Corresponding changes in federal policy, however, have been relatively limited. For example, the Employment Non-Discrimination Act (ENDA) has yet to pass both chambers of Congress despite majority public support in all states and nearly every House district (see our estimates in FIGURE 1). On other issues, there has been movement, but only with a long lag between opinion change and policy change. Since the high-profile murder of Matthew Shepard in 1998, large national majorities have favored treating violent acts against gays and lesbians as hate crimes; yet, Congress did not enact hate crimes legislation until Moreover, even when Congress has extended gay and lesbian rights, the margin of support typically has been much smaller in the legislature than among the public. To take a sneak peek at our findings later in this paper, consider the maps of opinion and voting on the 2010 House of Representatives vote to repeal Don t Ask, Don t Tell, ( DADT ) the policy banning gays and lesbians from serving openly in the military. Comparing the top left and bottom left panels in FIGURE 1, we see that while majorities in nearly all congressional districts outside the South supported repeal, only a much smaller majority of members of Congress voted for repeal. The 2007 House vote on ENDA, shown in the panels on the right, illustrates this pattern even more starkly. This is puzzling. Researchers have long demonstrated a strong correlation between public opinion and policymaking. This is one of the more robust findings in the empirical political science literature (Page and Shapiro 1983; Erikson, Wright and McIver 1993; Stimson, MacKuen, and Erikson 1995; Lax and Phillips 2012; inter alia). Furthermore, a variety of analyses suggest that elected officials face particularly strong incentives to respond to public sentiment on issues that are salient or related to morality policy (Mooney and Lee 1995, 2000; Haider-Markel and Meier 2003; Lax and Phillips 2009b, 2012). Gay and lesbian rights clearly meet both of these criteria. To be sure, this is not the only time that responsiveness to public opinion on high-salience civil rights issues has been limited. Examining early Gallup polls, Schickler (2010) finds that approx- 1

3 imately 60% of respondents supported anti-lynching legislation as early as 1937, while national majorities supported abolishing the the poll tax and enacting a federal law banning job discrimination in the 1940s. In spite of public opinion, Congress never passed an anti-lynching bill, and did not respond to demands for other types of civil rights legislation for two decades. Of course, this could result in part from what Katznelson et al. (1993) have called the Southern structural veto in Congress. Moreover, national majorities need not imply majorities in a majority of districts, as we see here with ENDA and DADT. So, does public opinion really drive policymaking in this area? The observations above suggest that the influence of public opinion may be more nuanced and limited than is often appreciated. To see how, we interrogate the relationship between public preferences and congressional action on gay and lesbian rights. We examine roll call voting on five major gay rights issues from the early 1990s to the present: same-sex marriage, adoption, hate crimes, employment non-discrimination, and military service. Using innovative roll call-specific estimates of public opinion, by state and congressional district, we evaluate not only whether higher constituent support makes legislators more likely to cast pro-gay votes, but also how often legislators vote in accordance with constituent opinion majorities, and how much public approval is necessary to stimulate support in Congress. On the surface, we find strong responsiveness to public opinion. We also find that a large majority (over two-thirds) of roll call votes are congruent with constituent opinion. This first cut is misleading, however. We show that this level of aggregation conceals troubling gaps in responsiveness and biases in policymaking as well as striking variation over time and between lawmakers. This more detailed picture of representation, broader and deeper than was previously possible, challenges and enriches our understanding of the opinion-vote relationship. 2 Studying Responsiveness Much of what we know about gay rights policymaking comes from the state level. Recent statelevel analysis (Lax and Phillips 2009b) shows that policy is generally responsive to opinion, in 2

4 that states with higher support for gay rights are more likely to enact pro-gay policies, but there is variation across policies when we look more closely. Furthermore, there is a surprising amount of non-congruence between opinion majorities and policy. For some policies, even clear supermajority support seems insufficient for policy adoption. When non-congruent, policy tends to be more conservative than desired by voters. While this analysis operates at a lower level of aggregation than prior work, which compared mood or ideology to policy, it does not reach the opinionvote level. Rather, it compares state-level opinion estimates to policy outcomes (i.e. whether the state adopted a given policy). This degree of aggregation makes it difficult to examine the extent to which constituent opinion drives roll-call voting, or how individual legislator characteristics shape responsiveness to public opinion. Existing national-level studies on gay rights also tend to operate at too high a level of aggregation to answer the questions posed above. They often aggregate policy areas, using indices for their dependent and/or key independent variables (Haider-Markel 1999, 2001; Lewis and Edelson 2000). Regressing indices of pro-gay roll call voting behavior on indices of pro-gay opinion masks cross-issue variation in opinion and responsiveness. This is problematic because public opinion polls consistently demonstrate that there is greater support for some pro-gay policies (ex. employment non-discrimination and hate crimes laws) than others (ex. same-sex marriage), and prior work shows differential responsiveness across issues. Indices of opinion and roll call votes also lack a common metric, without which it is impossible to determine whether lawmakers are voting in line with majority opinion in their district or state (Matsusaka 2001). Our analysis builds on existing work by extending an estimation technique to calculate issuespecific opinion by state and congressional district. We can then connect these estimates of constituent preferences to the corresponding representatives in Congress. This allows us to perform the first dyadic study of the opinion-vote relationship. When analyzing roll call votes on an employment non-discrimination bill, for example, we use estimates of district- and state-level support for adding sexual orientation to existing non-discrimination laws around the time of the vote. This allows both an individual-level and a system-level analysis. We can also contextualize the effects 3

5 of opinion, taking into account other influences on roll call voting and exploring differences between lawmakers. Finally, we offer not just a snapshot of the opinion-policy relationship, but an exploration of how it has changed over time. Examining the relationship between opinion and roll call voting on gay rights at the national level, what should we expect to find? We anticipate a strong opinion-policy relationship at the national/congressman-specific level that is similar to that previously found at the state system level, but we suspect this relationship will vary across different contexts and legislator types. Members of Congress have several sets of preferences to consider: their own, their party s, those of relevant interest groups, and those of their constituents. All legislators will not necessarily balance these potentially competing preferences in the same way some may be sensitive to constituent opinion, while others may be relatively impervious. We anticipate that growing public support for gay and lesbian rights will have predictable, differential effects across lawmaker types. First and foremost, we expect to find differences by party that grow over time. Since the early 1990s, the Democratic and Republican parties have staked out increasingly opposing positions on gay rights issues. Interest groups have also worked to push the parties to very different positions. The national party platforms first included references to gay and lesbian rights in 1992, with Democrats promising to provide civil rights protection for gay men and lesbians and an end to Defense Department discrimination and Republicans opposing any civil protections for gays and lesbians, as well as same-sex marriage and adoption. By 2008, the Democrats explicitly opposed the Defense of Marriage Act, supported the full inclusion of all families, including same-sex couples, in the life of our nation with equal responsibility, benefits, and protections, and promised to enact a comprehensive bipartisan employment non-discrimination act. They now, in 2012, support same-sex marriage. The Republican platform has stayed the same as it was or become even more explicitly anti-gay. There are now clear party positions on these issues, with pressure to adhere to them, even in the face of changing public opinion. Hence, Republican members of Congress may be less likely than their Democratic colleagues to respond to majority opinion. We might also expect less support for gay rights from Republican lawmakers, since members of 4

6 Congress might respond more to their fellow partisans than to the median voter, and Republican voters support gay rights less than Democratic voters. Next, lawmakers belonging to groups that have historically faced discrimination might be especially supportive of gay rights. Our analysis considers three such groups: women, Latinos, and African-Americans. Legislators who have faced discrimination, especially those who have been active in fighting against it, might be especially likely to protect minority rights, even against the wishes of their constituents. Such legislators might draw analogies between the civil rights struggles of their own groups and the fight for gay rights today. 1 They might be particularly aware of the need to set aside public opinion where minority rights are involved. While this is not representation in the direct majoritarian sense, it does fall in line with James Madison s argument, articulated most famously in Federalist 10, that one of the virtues of representative government is its ability to protect minorities from oppression by majorities. 3 Data and Methods Roll Call Votes. We evaluate the opinion-vote relationship in five issue areas, on a total of 23 separate roll call votes. Some involve bills proposing extensions of gay rights relative to the status quo, while others are on bills proposing conservative changes to the status quo. For each vote, our estimates of opinion are specific to the issue being considered, and rely on survey data collected around the time of (almost always before) the roll call vote. So, if we are considering a 2010 vote to repeal the military s ban on service by openly gay men and women, we use survey data from 1 We can see this kind of solidarity clearly in organizations advancing the rights of women and African-Americans. The country s largest and oldest civil rights organization, the National Association for the Advancement of Colored People (NAACP), has supported gay rights since the debate over open military service reached the national stage in Similarly, the National Organization for Women (NOW) has supported the cause since 1971, when it expanded its mission to include lesbian rights. Latino organizations have been less active on gay rights issues, and have not spoken out in favor of gay rights to the same degree as the NAACP and NOW. 5

7 2010 and Tables in the appendix list the roll call votes used and the survey data/questions employed to estimate state and congressional district opinion. Adoption: We have identified two votes in the House of Representatives, which occurred in 1998 and Both were on amendments to the Washington D.C. appropriations bill, seeking to prohibit couples in Washington D.C. who are not related by blood or marriage from adopting a child. The amendment passed in 1998, but was narrowly defeated in Same-Sex Marriage: There were three proposals. The first was the Defense of Marriage Act (DOMA), which defined marriage as a legally recognized union between one man and one woman. Under DOMA, the federal government would be prohibited from recognizing same-sex marriages and no state would be required to recognize same-sex marriages performed outside that state. DOMA passed the House and Senate in 1996 by wide margins. The second was the Federal Marriage Amendment (FMA), which sought to amend the U.S. Constitution to define marriage as a union between one man and one woman. The House and Senate voted on this proposal in 2004 and 2006, but the measure did not receive the requisite supermajority in the House and failed cloture votes in the Senate. The third proposal was an amendment to the Health Care and Education Reconciliation Act, which sought to suspend the issuance of marriage licenses to same-sex couples in the District of Columbia and require a public referendum. The Senate rejected this amendment in a 2010 vote. Gays in the Military: We have identified seven roll call votes pertaining to gays in the military, four of which occurred in 1993 and three of which occurred in Of the 1993 votes, two attempted to codify a full ban on military service by gays and lesbians, and two aimed to allow the President to decide the issue. We interpret a yes vote on the latter two amendments as a vote to allow gays to serve openly in the military, since this was the preferred position of President Clinton. All four votes failed. In 2010, the House voted twice and the Senate once on proposals to repeal DADT, the policy prohibiting the military from asking recruits about their sexual orientation, but allowing the military to discharge gay service members. Both proposals to repeal DADT passed. Jobs: We have identified three votes on employment non-discrimination. Two, a 1996 Senate 6

8 vote and a 2007 House vote, were on the Employment Non-Discrimination Act (ENDA), which sought to prohibit discrimination in employment on the basis of sexual orientation. In the Senate, it was defeated by a single vote. It then passed the House in 2007, without reaching a vote in the Senate. The third vote was a failed attempt by the House in 1998 to defund President Clinton s executive orders prohibiting discrimination in the federal civilian workforce. Hate Crimes: We have identified one vote in each chamber in 2000 and 2009 on a proposal to extend existing hate crimes protections to include sexual orientation. These proposals passed each time. In 2000, efforts to extend hate crimes protections died in conference committee. Opinion: MRP Goes to Washington. To calculate accurate estimates of public opinion for each of the 23 roll call votes in our analysis, we use multilevel regression and poststratification (MRP). This technique, first presented by Gelman and Little (1997), uses national surveys and advances in Bayesian statistics and multilevel modeling to generate opinion estimates by demographicgeographic subgroups. MRP has been shown to produce accurate estimates of public opinion by state and congressional district (Park, Gelman, and Bafumi 2006; Lax and Phillips 2009a; Rodden and Warshaw 2010), using as few data as in a single national survey and fairly simple demographicgeographic models (Lax and Phillips 2009a). Special limitations of standard survey data, however, necessitate an extension of the method, explained below. MRP proceeds in two stages. In the first stage, a multilevel model of individual survey response is estimated, with opinion modeled as a function of a respondent s demographic and geographic characteristics. We model survey response as a function of race, gender, age, education, state, region, state and congressional district presidential vote share, the percentage of the state/district that is African-American, aggregate state religious conservatism, and poll. These build on the standard predictors of social attitudes in general, and concerning gay rights in particular (e.g., Cook 1999). We find that demographic and geographic predictors perform quite well in explaining responses at the individual level. The results of this modeling stage are used to generate an estimate of opinion for each demographic-geographic type of voter. The second step is poststratification: the opin- 7

9 ion estimates for each demographic-geographic respondent type are weighted (poststratified) by the percentages of each type in actual populations of the relevant geography (either the state or congressional district). This allows us to estimate the percentage of respondents within each type who have a particular issue position or preference. Poststratification is done using state or congressional district population frequencies obtained from either the Public Use Micro Data Samples supplied by the Census Bureau or similar data. We convert the Census Public Use Microdata Area (PUMAs) to congressional districts using the Missouri Census Data Center s Geographic Correspondence Engine (geocorr2k) for each of the district boundary sets required by our analysis. For full details and validation of the basic method, see Lax and Phillips (2009a, b) for MRP by state and Rodden and Warshaw (2010) for MRP by district. One challenge remains in connecting the contributions of these works to actual substantive application of MRP at the district level. The survey data often do not include district identifiers, preventing direct use of district-level predictors in the modeling stage even if we have them at the poststratification stage. We have developed an extension of the MRP technique to allow researchers to create accurate district-level MRP estimates even where the survey data do not include these identifiers. This allows MRP at the district level for many more surveys than would otherwise be possible. Our modification to the basic MRP technique makes use of district-level adjustments that incorporate state-level correlations (which can be calculated even when district-level correlations cannot) and additional district-level information. For policies that have polls with congressional district identifiers, we use the coefficient on congressional district presidential vote share from the response model in the prediction stage for poststratification. For polls without congressional district identifiers, we use the coefficient on state presidential vote share from the response model (in lieu of a district-level coefficient) in combination with the congressional district vote share for prediction and poststratification. We do the same for share black in district. This modification is not only innocuous, as shown in FIGURE 2, but strongly improves estimate accuracy compared to using only state-level information throughout (details are given in the caption to the figure). Either way, the precise demographic makeup of each congressional district is used in the poststratifica- 8

10 tion, so that estimates would vary even without the adjustment we recommend. We use distinct poststratification files for the period before and after the 2000 national redistricting, after the 2003 Texas redistricting, and after the court-required 2003 Texas redistricting. TABLE 1 shows summary statistics of issue-specific opinion around the time of each roll call vote, with histograms thereof in FIGURE 3. Across all votes, our opinion estimates are coded in the pro-gay direction, such that higher values always indicate more pro-gay opinion. There is a large range in public opinion across states and congressional districts, and there is also variation in the percentage of districts with majority support across issues (DOMA had no pro-gay majority states/districts in 1996, while hate crimes in 2009 had only pro-gay states/districts). Overall, mean pro-gay opinion is over 50% for 14 of the 23 roll calls. The roll call vote with the highest support for the liberal position, at an average of 73%, was the 2007 House vote on ENDA. Pro-gay opinion was lowest on the Senate and House DOMA votes in 1996, averaging just below 30%. 4 Basic Responsiveness and Congruence Responsiveness. We begin by considering responsiveness at the level of individual roll call votes. Each graph in FIGURE 4 takes one roll call vote and plots the probability of an individual legislator casting a pro-gay vote on the y-axis, against our estimates of their state or congressional district s opinion on the x-axis. The small tick marks along the top (bottom) axis indicate the opinion in states or congressional districts whose representative cast a pro-gay (anti-gay) vote, and the bold line shows the logit regression curve. All of the curves indicate a statistically significant responsiveness coefficient (at the 95% level), meaning that the probability of a lawmaker casting a pro-gay vote is positively correlated to policy-specific opinion. The gray line is a lowess curve, which shows that the vote-opinion relationship does not require assuming a logistic functional form. Responsiveness is strong on a policy if the curve is steep and positively sloped. The strength of the opinion-vote relationship varies across policies, as indicated by differences in the slopes of the logit curves. For example, the logit curve for the 2006 vote in both the House 9

11 and Senate on the Federal Marriage Amendment is much steeper than the 1993 Senate votes to codify the existing ban on gays and lesbians serving in the military ( DADT1993bansenate ). In general, however, most of our logit curves are upward sloping and steep, indicating a strong overall correlation between opinion and policy. Congruence. If we look at congruence with majority opinion, however, the opinion-vote relationship appears less strong. Indeed, roll call voting often has a conservative or liberal bias. To see this, consider the location of the logit curves, which can be shifted left or right of center, meaning either less or more than 50% support is needed, respectively, to stimulate a pro-gay roll call vote. The dotted lines extending from the x- and y-axes indicate where public support and the probability of a pro-gay vote each reach 50%. The point at which the regression curve intersects the vertical dotted line reveals (read on the y-axis) the predicted probability of a pro-gay roll call vote when public support is 50%. The point at which the regression curve intersects the horizontal dotted line reveals (read on the x-axis) the needed level of support for the predicted probability of a pro-gay vote to reach 50%. For perfect majoritarian responsiveness, the slope of the estimated logit curve would be very steep at 50% and pass through the crosshair within each panel, since the crosshair marks the point at which 50% support leads to a 50% probability of casting a pro-gay vote. For some votes (ex. FMA2006senate ), the curve comes close to this majoritarian ideal. If the responsiveness curve goes through the crosshair, but is insufficiently steep, we have policy bias, but this bias is ideologically balanced (i.e., some is conservative, some is liberal, and they cancel out). However, if the curve is shifted to the right of the crosshair, then a supermajority of liberal opinion is needed to bring about a pro-gay vote. This indicates a conservative bias in policymaking. A curve shifted to the left of the crosshair demonstrates the opposite. The farther the logit curve falls from the crosshair, the greater the magnitude of the bias. Even a cursory glance at FIGURE 4 shows that we do not always observe perfect majoritarianism. Indeed, our results include clear examples of strong responsiveness (a steep curve) coinciding with frequent mismatches between majority will and policy, due to biased policy curves. The

12 House vote on employment non-discrimination ( JOBS2007house ) and the House 2009 vote on hate crimes ( Hate2009house ) both show a significant positive relationship between opinion and a lawmaker s vote and have steep responsiveness curves. When plotted, however, both fall to the right of the crosshair. This shift means that support has to be far above 50% to have a 50% chance of a pro-gay roll call vote. This helps to explain the discrepancies between opinion and voting displayed in the maps in FIGURE 1. Indeed, for JOBS2007house, constituent opinion needs to be 71% before the lawmaker has a 50% probability of casting a liberal vote. For Hate2009house constituent opinion needs to be 68%. The consequence is that congruence for both is relatively low only 56% for JOBS2007house and 59% for Hate2009house. In comparison, congruence for the 2006 Senate vote on the Federal Marriage Amendment (which has a responsiveness curve that passes through the cross hair) is a whopping 80%. Overall, we find that 68% of the 4,982 terminal roll call votes included in our analysis are congruent with majority opinion. If we only consider roll call votes where the size of the opinion majority is greater than 60%, then congruence rises to 78%. For majorities of size 70% or more, it reaches 86%. Of course, congruence varies substantially by vote. The percentages listed in the fourth column of TABLE 1 are the shares of roll call votes that match the district or state opinion majority. Across roll calls, congruence ranges from 56% ( Hate2000house and JOBS2007house ) to 86% ( DOMA1996senate ). Votes most frequently match majority opinion on bills dealing with same-sex marriage (74%), while they tend to be least congruent on hate crimes bills (61%). To put the 68% overall congruence figure for congressional votes into perspective, Lax and Phillips (2009b) find a similar 62% level of congruence between opinion and policy (not votes) at the state level, which is significantly higher than the 48% congruence level Lax and Phillips (2011) find over a much larger set of policy types. Matsusaka (2010) finds a 57% congruence level for a subset of those policies. Finally, Monroe (1998) finds a 55% match between national policies and national opinion majorities over a wide set of issues. Gay rights majoritarianism does not appear to be very different at the state and national levels. This issue area also does not appear to demonstrate below-average congruence levels; if anything, there seems to be more 11

13 congruence on gay rights than on other issues. The ideological direction of incongruence varies across groups. Incongruent votes cast by Republicans are, unsurprisingly, only liberal 6% of the time, compared to 76% for Democrats. Breaking this down by race, we can see that 74% of incongruent votes cast by white Democrats are liberal, compared to 85% for black democrats. Facing a pro-gay opinion majority, black lawmakers vote pro-gay 95% of the time; facing an anti-gay opinion majority, they still do so 75% of the time. Compare this to white lawmakers, who, facing a liberal opinion majority, cast a pro-gay vote 58% of the time. Facing a conservative opinion majority, they cast a pro-gay vote 22% of the time. For white Democrats, these numbers are 93% and 48%; for Republicans, they are 19% and 7%. Also consistent with the aforementioned state-level findings, we observe far more conservative bias than liberal bias. Of the 2,089 incongruent roll call votes in our dataset, only 552 (26%) are in the liberal direction. In other words, when lawmakers vote against the preferences of their constituents, they tend to take a more conservative position. Another way of describing this is that liberal opinion majorities get what they want 63% of the time, while conservative opinion majorities prevail 75% of the time. TABLE 1 summarizes the ideological tendency of incongruence for each roll call vote. The fifth column shows the share of incongruent votes that are liberal. For only 7 of 23 roll call votes do we see more liberal than conservative incongruence. Of course, total incongruence could be small or incongruence of different types could cancel out. So, we also calculate the net number of pro-gay votes gained or lost due to incongruence. The sixth column shows the net liberal vote bias the number of liberal incongruent votes minus the number of conservative ones. This measure demonstrates that anti-majoritarian voting behavior tends to be associated with a deficit in support for gay rights in Congress. In the House, for example, the greatest benefit the pro-gay side ever received from incongruence amounted to 79 votes ( DADT1993hunterhouse ), while they lost more than 150 votes four times (see, for example, HATE2009house where the pro-gay side lost 172 votes). Without incongruence (i.e. if pro-gay voter majorities won out), four roll call votes would have flipped into pro-gay simple majorities: FMA2004house, FMA2006house, FMA2006senate, and JOBS1995senate. On the other 12

14 side, three roll call votes had pro-gay legislative majorities that would not have existed if all votes had been congruent with opinion majorities: DADT1993senate, DADT1993hunterhouse, and DCMARRIAGE2010senate. This does not necessarily mean the policy outcomes would have changed, since simple majorities are not filibuster-proof. Overall, Democrats are congruent 82% of the time, while Republicans are congruent 52% of the time. Here, there is only a small difference between white and black Democrats (82 versus 80, respectively). Now, we turn to an analysis of public opinion s influence on voting in context. 5 Opinion in Context: Other Influences on Roll Call Voting We show multivariate models of responsiveness in which the dependent variable indicates whether the roll call vote cast was pro-gay (liberal). We combine all votes into a single mega model. The variables include the following: Opinion: percentage of constituents (of those with an opinion) who support the pro-gay position on the roll call vote Republican: coded one if the lawmaker is a Republican, zero otherwise DW Nominate 1: liberalness of the lawmaker on the first dimension identified by Poole & Rosenthal, based on her roll call voting behavior DW Nominate 2: liberalness of the lawmaker on the second dimension identified by Poole & Rosenthal, based on her roll call voting behavior Male: coded one if the lawmaker is male, zero otherwise Latino: coded one if the lawmaker is Latino, zero otherwise White: coded one if the lawmaker is white or Asian, zero otherwise Senate: coded one if the lawmaker serves in the Senate, zero otherwise TABLE 2 displays results from eight model variants, to check robustness across specifications and to facilitate various all else equal comparisons (which is also to say that caution must be taken in interpreting raw coefficients). We allow for varying intercepts and slopes for opinion. 2 2 Key results are unchanged if we use additional random effects by state or member of Congress, 13

15 The basic relationship between voting and opinion holds: lawmakers whose constituents have higher levels of pro-gay opinion are far more likely to cast pro-gay votes. This effect holds even after controlling for a lawmaker s party and ideology, which are themselves strongly influenced by generalized constituent preferences. The effects of opinion also remain when we control for Democratic presidential vote share in the state or district, and other similar predictors. At an average value of opinion (in Model 1), one additional point of policy-specific opinion increases the chance of policy adoption by approximately 5 percentage points. Party is also a clear predictor of voting (e.g., Models 2, 4, and 7; in Models 6 and 8, the coefficient on Republican captures the effect thereof after controlling for Nominate score, making for a strange all-else-equal comparison). Model 4 also shows that blacks and Latinos tend to vote pro-gay relative to whites, controlling for opinion (and not controlling for Nominate). Models 5 and 6 show almost no difference between blacks and whites once we control for Nominate, but this is only true on average as the discussion below will reveal. Female members of Congress vote more pro-gay than their other predictors would suggest, no matter which model we examine (that is, the coefficient on female is always positive and significant, statistically and substantively). 6 Deeper Results Analyzing differences by party, gender, and race provides an even more nuanced view of responsiveness. Models 7 and 8 allow for the effects of opinion to vary by congressman type. We summarize these differences using Model 7 to calculate the level of pro-gay opinion needed for a 50% probability of casting a pro-gay roll call vote for different types of legislators, ordered from most pro-gay to least (for the average roll call): Black Female Democrats: 31% Black Male Democrats: 38% or fixed effects for state or issue. Results are similar if we shift the dependent variable to congruence. Results are similar if we allow the slope of DW Nominate to vary by demographic subgroup. We find very little difference between the House and Senate in responsiveness to opinion. 14

16 White Female Democrats: 40% Latino Male Democrats: 44% White Male Democrats: 46% White Female Republicans: 58% White Male Republicans: 66% Women are more likely to cast pro-gay votes, compared to co-partisan men. However, gender has a much less dramatic effect than party and race. While Latino lawmakers are similar to white lawmakers, black lawmakers are substantially more supportive of gay rights than their white colleagues. As do the raw congruence numbers, these numbers indicate that the ideological direction of bias varies by legislator type. Not only are there differences in the level of constituent support necessary for legislators to cast pro-gay votes, but the effect of increasing opinion also differs across groups. For each additional point of policy-specific opinion (based on Model 7), the probability of a white male Democrat casting a pro-gay vote rises by 5. For white Republicans, the probability only rises by 4, and for black Democrats, it only rises by 3. These differences reflect the significant interactions between opinion and party, and opinion and race, shown in Model 7. The interaction of opinion and gender was not significant. We can see the same patterns by looking at the regression results, as well as the raw data in Figure 5. Each graph plots mean pro-gay opinion against the percentage of pro-gay votes cast by each legislator. The size of the points shows the number of votes represented (from 1 to 14). Republicans are marked with squares, white Democrats with circles, and black Democrats with triangles. The dashed lines are lowess curves. The top-left panel captures the positive overall relationship between opinion and roll call voting, captured in models 1 through 6. However, the remaining three panels present clear evidence of differential responsiveness by party and race. The top right graph shows opinion and voting by black Democrats. As the flat lowess curve suggests, there is a much weaker relationship between opinion and voting for black Democrats than for non-black Democrats (see the bottom left graph). Black members of Congress are concentrated at the top of 15

17 the graph over 70% support gay rights 100% of the time. Regression results confirm that black legislators are more likely to cast pro-gay votes than white legislators (see the positive, significant coefficient on black in Model 4). Additional pro-gay support matters less for black legislators than white legislators (as indicated by the negative and significant coefficient on the interaction with opinion in models 7, 8, and 9). Looking at model 7, the slope is 6.0 for white male Democrats (see the coefficient on opinion), 5.2 for Republicans (see the coefficient on Republican*opinion), and 4.0 for blacks (see the coefficient on black*opinion). Looking back at Figure 5, the raw data suggests that Republican congressmen are also less responsive to opinion than Democrats. Two-thirds of Republicans in our sample have never cast a pro-gay roll call vote, regardless of opinion. Regression results confirm that Republicans are less likely to cast pro-gay votes than Democrats (see the negative, significant coefficient on Republican in model 4), and also less responsive to opinion (see the negative, significant coefficient on the interaction between Republican and opinion in Models 7, 8, and 9). 3 In sum, Figure 5 makes clear that the overall direct relationship between voting and opinion (in the top left panel) is generally due to Democratic non-black congressmen, and that neither Democratic black congressmen nor Republicans show much of a relationship between opinion and voting. Indeed, white male Democrats need 46% support to have a 50% chance of voting pro-gay. This is the closest to a match of any subgroup. We can see, in contrast, the extent to which ideology predicts pro-gay voting by looking at Figure 6. This is identical to the previous figure, except that the x-axis represents mean Nominate score (liberalism) instead of opinion. The top left graph demonstrates a strong relationship between ideology and roll call voting. This is consistent with regression results, which show that DW Nomi- 3 Party seems to perform inconsistently across models. However, it is not that the effect of party varies so greatly across specifications. Rather, these models each have a different conception of holding all else equal. Which model one should use to evaluate the impact of being Republican depends on what the substantive questions is that is, what one wishes to hold constant. (It can be a rather odd thought experiment to hold constant the Nominate score of a black Democratic congressman when changing him to a white Republican congressman.) 16

18 nate is always a positive and significant determinant of voting behavior (i.e. more liberal members of Congress are more likely to cast a pro-gay roll call vote). Even controlling for differences in the slopes of Nominate across groups, the opinion relationships remain. Of the three subgroup graphs, the bottom left (for Republicans) shows the strongest relationship between ideology and voting. Comparing this graph to the bottom left graph of Figure 5 suggests that Republicans are voting according to their ideology, rather than opinion. Figure 7 facilitates a more direct comparison between opinion and ideology. Each circle represents a member of Congress, plotted by mean pro-gay opinion (across issues) and Nominate score. The size of the points indicates the number of votes represented (from 1 to 14). The shading captures how pro-gay the voting record was, ranging from light gray (0%) to black (100%). The dashed line shows where the voting record is predicted to be 50% pro-gay, based on OLS regression on opinion and Nominate score, weighted by number of votes cast. Dotted lines to the left and right of it show the 25% and 75% levels respectively. Note that if voting were only responsive to legislator ideology (not opinion at all), the iso-lines would be vertical. If only opinion mattered, and not legislator ideology, the iso-lines would be horizontal. The lines are steepest for black Democrats, showing how little opinion matters for their votes, and least steep for non-black Democrats, for whom opinion matters more than for Republicans. Overall, we find strong evidence for our hypothesis that support for gay rights should be especially high amongst legislators belonging to groups that have historically faced discrimination. Though our findings about Latinos depend on model specification, we consistently find that African-American and female representatives are especially likely to cast pro-gay votes. Turning to differential responsiveness, however, race seems to be more influential than gender. While women are not any more or less responsive to growing support for gay rights than men, changes in opinion have less influence on African-American legislators than their white colleagues. This is primarily because black legislators strongly support gay rights, even if their constituents do not. To some extent, the patterns noted thus far form a snapshot. We now show that disaggregating 17

19 by party over time reveals important details. 4 Vote Switching. One of the benefits of analyzing responsiveness over a long time period is that we can study lawmakers who cast multiple votes on the same issues over time, assessing whether opinion change matters (we might not expect votes to change if opinion has not). There are five issues with roll call votes in different sessions: adoption (1998, 1999), Don t Ask, Don t Tell (1993, 2010), the Federal Marriage Amendment (2004, 2006), hate crimes (2000, 2009), and employment non-discrimination (1998, 2007). Since support for gay rights has been steadily increasing over time, we are most interested in studying the extent to which lawmakers shifted from opposition to support for gay rights. Hence, we will focus on the 687 (of 1,453) pairs in which the first vote was against gay rights (note that this is not a random sample of lawmakers, which should be kept in mind when interpreting the following findings). Overall, vote switching is rare, occurring in only 6.3% of vote pairs (91 switches). However, when it does occur, it is correlated with opinion change and party. When opinion change was positive, 13% switched. When opinion change was negative, under 4% switched. Flipping this, switchers had seen an 8% increase in pro-gay opinion between votes on average, while nonswitchers had seen only a 1% increase. On the partisan dimension, Democrats are much more likely than Republicans to switch. White male democrats who started out with an anti-gay vote in an anti-gay district, whose district shifted to being pro-gay, had a 65% chance of switching to the pro-gay position in the second vote. In contrast, white male Republicans whose districts started out 4 Other results: We also compared terminal to procedural votes, finding that 68% of the former were congruent, compared to 60% of the latter. This 8% gap did not have a specific ideological direction, however some was liberal, some was conservative, and it cancelled out. If we examine salience, we find no difference in congruence between roll call votes above and below the average level of salience. This whole issue area is of high salience, however. It is also not the case that legislators in more marginal districts are more congruent with majority opinion. We did find, though, that those who were marginal in their general election were very slightly less responsive to opinion, while those who were marginal in the primary were very slightly more responsive to opinion. Overall, however, marginality does not seem to play a major role. 18

20 as anti-gay and switched to pro-gay had only a 4% chance of switching. Where opinion change was positive, over half of Democrats switched, but only 3% of Republicans did. (It is difficult to draw inferences relating to race and gender because almost all female and minority legislators began with pro-gay votes.) FIGURE 8 plots pro-gay constituent opinion at the time of the first vote against pro-gay opinion at the time of the second vote. Each point represents a legislator who voted twice on a particular issue. Since there are five issues, each legislator could theoretically be on the graph up to five times. Points to the right of the vertical dotted line (representing the 50% majority threshold for time one) represent legislators whose first votes were incongruent with majority opinion. 350 legislators, or slightly more than half of the sample, fall into this category. The key area, from a responsiveness standpoint, is the top left quadrant. Here, we have lawmakers whose constituents did not support gay rights at the time of the first vote, but did at the time of the second. In a world of perfect majoritarian responsiveness, all of the dots in this quadrant would be black (i.e. every legislator should switch his vote). However, this is far from the case. In reality, only 16 of the 78 legislators (21%) in this quadrant switched their votes. There is a clear partisan dimension to this trend, as 14 of 23 Democrats switched, while only 2 of 55 Republicans did the same. We are not seeing fewer Republican switchers simply because their constituents are conservative on gay rights. As this graph indicates, many Republican states and districts crossed the 50% threshold between the two votes; yet, their representatives did not follow. In the top right quadrant, we could see corrective votes: legislators who were incongruent at time one could become congruent at time two. This is rare. Only 22 of the 272 legislators (8%) in this quadrant switched their votes. The remaining 250 legislators voted against gay rights in both sessions, against the wishes of their constituents. The vast majority of legislators 247 of 272 in this quadrant are Republicans; and of these, only 3% switched to the pro-gay position in the second vote. In contrast, over half of the Democrats in this quadrant 13 of 25 switched. Given lawmakers reticence to change their votes, even in the face of opinion change, it is puzzling that 17 of 259 legislators whose constituents did not support gay rights at either time switched 19

21 to the pro-gay position in the second vote. The 12 Democrats (of 53) in the lower left quadrant may have been following their party s position, which liberalized over time. The 5 Republican switchers seem odd, but represent only 2% of Republicans in that quadrant. Opinion and Voting over Time. Setting aside party again, we assess time trends in opinion and voting, and the relationship between them. FIGURE 9 plots these over time. A data point, plotted with its substantive label, represents the average over districts or states, as appropriate. The dashed line shows the average trend (lowess curve), and the dotted lines shows trends by issue. While there is some variation in the time trends by issue, we focus on the overall trends. Reading these panels in order tells the following story: [1] Mean pro-gay opinion has increased over time, from around 45% to around 60%. [2] The number of pro-gay opinion majorities has increased far more sharply, from around 35% to 85%. [3] However, the percentage of pro-gay roll call votes cast has increased far less dramatically (from 50% to 60%). [4] Surprisingly for now overall congruence is nearly constant (around 70%). We will soon see why. [5] and [6] On the other hand, the nature of incongruence has changed drastically. Incongruence, once leaning to the liberal side, now strongly cuts against pro-gay policy, measured either as a percentage of incongruence or by the net vote bias (under +15% to -25%). The predicted conservative vote bias from incongruence now averages 109 votes in the House (i.e. 109 votes are lost because lawmakers are not following constituent opinion) and 25 in the Senate. This shows significant change over time, missed by our earlier snapshot, in which we showed that the pro-gay side lost an average of 73 votes in the House and 12 votes in the Senate (see TABLE 1). That snapshot also obscured the liberal voting bias occurring early in our time period, which cancelled out in the mid-1990s before tilting conservative. We next show what can be learned by breaking this down further, by party. Party. Consider the patterns of opinion and voting by party illustrated in Figure 10. Five of the panels correspond to the panels in the previous figure and are labeled to indicate that. [1P, breaking down Figure 9, Panel 1] Support for gay rights has grown in both Democratic and Republican 20

22 constituencies over time. While the parties started roughly in the same place, in terms of opinion in their districts/states, Democrat-represented constituencies (DRCs) have grown more liberal at a slightly higher rate than Republican-represented constituencies (RRCs), leading to a small party gap today (roughly 5 percentage points). [2P] There is a much larger gap between the parties in terms of opinion majorities, but this gap has stayed constant over time. DRCs went from being roughly split between pro-gay and anti-gay majorities to nearly 100% pro-gay majorities. However, even a majority of RRCs have been majority pro-gay since the late 1990s (75% as of 2011). Within Congress, however, we observe something different. [3P] Democrats have steadily voted more pro-gay over time, starting from a relatively high base rate, moving towards near unanimity. Republicans, however, have remained relatively constant around a much lower rate (less than 15%) and have not changed along with their RRCs. In sum, as Democrats have moved into sync with their constituents, Republicans have dropped out of sync with theirs. [4P] Since RRC pro-gay majorities have become far more common, Republican congruence rates have plummeted from 75% to 35%. Meanwhile, Democratic congruence has increased. [6P] This is partly because DRCs have moved in line with the initial pro-gay voting rates observed, and partly because Democrats have moved to match their increasingly pro-gay DRCs by voting along those lines. The initial liberal vote bias by Democrats disappeared by the early 2000s, and they have stayed in line on average since (that is, the remaining incongruence cancels out). On the other hand, Republicans, who started out with balanced incongruence, now show clear conservative bias relative to their RRCs. This breakdown by party shows that it is only the Democrats who have voted more pro-gay over time. Congruence remained constant overall because the increase in Democratic congruence made up for the decrease in Republican congruence. Until 2004 or so, both parties contributed to the growing conservative vote bias, with Democrats losing their liberal bias and Republicans gaining a conservative bias. Since then, Democrats have stayed in sync while Republicans have increased their conservative vote bias passively by not following their constituents as they tipped over into pro-gay majorities. 21