Abstract
Prior public opinion research has identified a wide range of circumstances in which polling results may be tainted by social desirability bias. In races pitting a Black candidate against White opponents, this has often been referred to as the “Bradley effect” (aka “Wilder effect” or “Dinkins effect”), by which survey respondents overstate their preference for Black candidates running against White opponents. This study examines the accuracy of polling on same-sex marriage ballot measures relative to polling on other statewide ballot issues in all states voting on the issue from 1998 to 2012, controlling for a range of theoretically relevant contextual factors. There has been a great deal of speculation, though little empirical evidence, that polling systematically understates opposition to same-sex marriage. Consistent with social desirability bias, this study finds that opposition to same-sex marriage is about 5% to 7% greater on election day than in preelection polls.
Keywords
Introduction
In the late 1980s, a number of political observers began to notice a pattern in which Black candidates running against White opponents seemed to fare much worse on election day than predicted by preelection polls. For example, in 1989 Douglas Wilder was elected as the first Black governor of Virginia and only the second Black governor of any state in U.S. history. Despite its historic significance, Wilder’s victory was widely anticipated at the time; in the weeks leading up to the election, polls showed him leading by 15 points or more (Melton & Morin, 1989). What was surprising was his razor-thin margin of victory over Marshall Coleman, a White candidate, by less than half a percent.
Wilder’s precipitous decline in the polls may have been chalked up to the vagaries of his specific race had it not been accompanied by similar occurrences in other races around the country. For example, in the same election cycle, David Dinkins, a Black candidate, won the mayoral race in New York city by beating Republican Rudolph Giuliani by less than 2% of the vote despite having led in some preelection polls by close to 20%. This led observers to recall the 1982 California gubernatorial election in which Tom Bradley, the Black mayor of Los Angeles, lost to George Deukmejian by a narrow margin after being heavily favored in the polls 1 .
Taken together, these races led to speculation that Black candidates, when facing a White opponent, tend to fare much worse in actual elections than in preelection polls. The assumption, rooted in the concept of social desirability bias, was that many poll respondents falsely voice support for Black candidates in an effort to appear racially tolerant. This pattern of responses became popularly known as the “Bradley effect,” “Wilder effect,” or “Dinkins effect. 2 ”
In recent years, anecdotal evidence suggests that a similar effect may be at work in another setting. Between 1998 and 2012, there were 33 state votes on ballot measures concerning same-sex marriage. Thirty of those ballot measures were designed to restrict marriage to one man and one woman—thus barring same-sex marriages—while the other three measures were votes to specifically legalize same-sex marriage. Three states (Arizona, California, and Maine) voted twice on the issue. Same-sex marriage bans have won voter approval in every instance except two (Arizona in 2006 and Minnesota in 2012), while three states (Maine, Maryland, and Washington) voted in 2012 to legalize same-sex marriage. In many of the states voting on the issue, observers noted that opposition to same-sex marriage appeared to be significantly higher at the ballot box than in preelection polls. The case of Maine’s 2009 vote is indicative of this pattern. In the final preelection poll, only 40% of Maine voters said they planned to vote in favor of Question 1, which would restrict marriages to opposite-sex couples. However, on election day Question 1 passed with 53 % of the vote. According to Tom McCluskey of the Family Research Council, a group that opposes same-sex marriage, “We’ve seen it, I think, in every single case, that it [opposition to same-sex marriage] is under-polled every single time. I’ve seen much higher, but normally we add 5 to 10 percentage points to any polling (Buchanan, 2006, p. A4).” According to Michael Traugott of the University of Michigan, “People are more likely to say they support individual or group rights even when in some cases they don’t (quoted in Buchanan, 2006, p. A4).”
Although there has been quite a bit of speculation about the existence of social desirability bias in voting on same-sex ballot measures, very little empirical research exists on the subject. This study is designed to fill this void by examining whether or not polling on such measures systematically understates opposition to same-sex marriage.
Social Desirability in Public Opinion Polling
Numerous studies of public opinion polling methods have found that poll results can be distorted due to a host of potential flaws in survey design and implementation. In terms of the Bradley effect, most of the speculation has centered on social desirability bias, the tendency of respondents to give answers they perceive to be socially desirable regardless of their own true positions (Aquilino, 1994; Bishop & Fisher, 1991; Finkel, Guterbock, & Borg, 1991; Fowler, Roman, & Di, 1998; Keeter & Samaranayake, 2007; Reeves, 1997; Streb, Burrell, Frederick, & Genovese, 2008; Traugott & Price, 1991).
Indeed, a long line of research indicates that social desirability bias can be a serious concern in a range of polling contexts. For example, measures of voter turnout based on respondent self-reporting are now subjected to careful verification procedures since people often have a tendency to say they voted in elections they didn’t (Katosh & Traugott, 1981; Presser, 1990; Silver, Anderson, & Abramson, 1986). Polling on issues and candidates with racial and gender dimensions presents particular issues for researchers. Studies have shown that respondents can be influenced by the perceived race, gender, and ethnicity of interviewers (Cotter, Cohen, & Coulter, 1982; Huddy et al., 1997; Kane & Macaulay, 1993; Kinder & Sanders, 1996; Reese, Danielson, Shomaker, Chang, & Hsu, 1986; Weeks & Moore, 1981) or the perceived racial dimensions of the candidates or issues being asked about (Berinsky, 1999, 2002; Janus, 2010; Krysan, 1998; Streb et al., 2008). Several studies have employed a novel approach known as a list experiment to gauge the extent to which social desirability is an issue in these sorts of polling contexts (Janus, 2010; Kane, Craig, & Wald, 2004; Kuklinski, Cobb, & Gilens, 1997; Sniderman & Carmines, 1997; Streb et al., 2008).
Of course, social desirability biases will vary depending upon the cultural context. For example, in the United States the public expression of racial bias against Blacks has become increasingly taboo in recent decades as racial tolerance has become a more widely accepted norm in public behavior. Given the checkered past of race relations in the United States, and more recent changes in peoples’ attitudes, social desirability biases have been hypothesized in polling on races pitting a Black candidate against a White opponent. In other words, some individuals who favor the White candidate will actually express support for the Black candidate in an apparent attempt to appear racially tolerant to the interviewer. Similarly, Berinsky (1999) finds that some respondents will claim to be undecided instead of voicing opposition to a Black candidate.
For many years, the existence of a Bradley effect survived in the realm of anecdotal evidence, speculation, and so-called conventional wisdom. Recently, Hopkins (2009) undertook a more rigorous empirical study of this effect by examining polling data and election results in all U.S. senatorial and gubernatorial races from 1989 to 2006 that pitted a Black candidate versus a White one. Controlling for a range of relevant contextual factors, Hopkins found the Bradley effect had been a significant factor in such races in the past, averaging about 2.6% points. However, Hopkins found the Bradley effect ceased to exist after 1996, although he is unable to provide a definitive explanation about why this is the case.
At the individual level, the key factors that affect the likelihood of a respondent being swayed by social desirability appear to be the perceived or assumed viewpoints of the interviewer and the extent to which an individual respondent is “self-monitoring” (Snyder, 1987). This makes it exceedingly difficult to sort out the theoretical linkages between aggregate measures of opinion on an issue and the prevalence of social desirability bias in that same domain. Hopkins’s (2009) study, discussed above, also presents us with difficulties for disentangling the relationship between aggregate trends at the national or state level and the prevalence of social desirability bias. Specifically, one would expect that such polling distortions on racial issues would increase as aggregate opinion about race liberalizes, because respondents would be more hesitant to voice opinions opposed by larger percentages of fellow citizens. This expectation is consistent with Noelle-Neumann’s (1974) finding of a “spiral of silence,” in which those who perceive themselves to hold opinions in a societal minority fail to speak out in fear of social ostracism. However, Hopkins finds the Bradley effect has decreased over time as there has been a significant, long-term liberalization in racial attitudes among Americans. At some point, it may be that social desirability diminishes when an overwhelming consensus emerges because there are simply fewer people in opposition to the dominant opinion. In any event, this is a puzzle that has not been fully solved in the literature thus far.
Although publicly expressed tolerance of homosexuality in the United States has not reached the same level as that on racial issues, Americans’ attitudes about homosexuality have been liberalizing in recent decades (Gallup, 2012a; Putnam & Campbell, 2010). According to Gallup (2012a), a majority of Americans have supported marriage rights for same-sex couples since 2011, following a general trend of increasing support over many years. Still, there are important differences between issues related to race and sexuality. Whereas very few Americans would still voice support for banning interracial marriage (Kinder & Sears, 1981; Mendelberg, 2001; Sears, Henry, & Kosterman, 2000; Sears, Hetts, Sidanius, & Bobo, 2000), only a narrow majority of Americans voiced support for same-sex marriage as of 2012 (Gallup, 2012a). In other words, while public attitudes about homosexuality, and the corresponding public policy issues, have been undergoing gradual, meaningful change, those changes still lag behind racial attitudes by quite a few years. As a result, we can hypothesize that social desirability bias may remain an issue in polling on same-sex marriage ballot measures for the foreseeable future.
It is also easy to hypothesize explanations other than aggregate opinion trends that might account for changes in the prevalence of social desirability bias in polling. For example, in the past several decades public education and cultural norms in the United States have increasingly stressed the need for people to be more sensitive in their awareness of diverse populations and the public language used to discuss such differences. Although such norms certainly vary in the extent of their adoption across different segments of the population, there is nevertheless a greater overall social stigma associated with the expression of potentially discriminatory, insensitive, or unfavorable opinions related to topics such as gender, race, ethnicity, disabilities, and sexuality.
For these reasons, sorting out the theoretical linkages between aggregate opinion measures and the prevalence of social desirability bias is a complicated task, mostly based on the fact that existing studies have primarily used aggregate data to study what is fundamentally an individual-level phenomenon. The purpose of this study is to examine the accuracy of polling on same-sex marriage ballot measures relative to other types of ballot questions at the state level. As such, the data and methods described in the following section are theoretically grounded in the existing literature on social desirability bias, while remaining cognizant of the difficulty in reaching definitive conclusions about the “black box” of individual decision-making processes using aggregate, state-level data.
Data and Method
In order to examine social desirability bias in the context of polling on ballot measures concerning same-sex marriage, I examined all statewide races from 1998 to 2012 in which voters were asked to vote on such measures and in which statewide polling was conducted prior to election day. Although there has been a great deal of polling done on the public’s general approval of same-sex marriage, polling on specific ballot initiatives has been less regular. In four of the state races, no such polls were conducted so they are excluded from the analysis.
Conveniently, in most states voting on same-sex ballot measures, there were other questions on the ballot at the same time. The issues covered by those other measures included a wide-range of issues that would not be expected to generate social desirability biases—taxes, bond issues, term limits for state legislators, etc. Thus, I collected polling data from those races as well, providing a valuable control for a range of potential polling errors, house effects, and other distortions. Moreover, much of the analysis presented below focuses on those instances in which respondents were questioned about same-sex marriage and other ballot measures in the same survey conducted by the same polling firm, thereby controlling for a whole host of other potential polling biases.
The polls examined in this study, for same-sex and nonsame-sex ballot measures, were identified through two primary sources. First, I consulted the iPoll Databank maintained by the Roper Center at the University of Connecticut. Even though iPoll is generally regarded as the most comprehensive archive of public opinion data, it is incomplete, particularly for state-level polls and those conducted by commercial polling firms. Therefore, I also performed comprehensive searches of Lexis-Nexis in order to identify all relevant polls referenced in the media. For the purposes of this study, I included all polls that were conducted within 90 days of each election. Thus, the polls examined here include all of those publicly available, not a sample. The dataset includes 59 polls on same-sex marriage ballot measures and 77 polls on ballot measures regarding other issues. For each poll, additional information was collected on sample size, margin of error, sampling method (registered vs. likely voters), and interviewer type (live vs. automated interviewer). Thus, the unit of analysis in the models described below is each distinct poll on a state ballot measure, including both same-sex marriage and nonsame-sex marriage issues.
Since social desirability bias relates to the gap between polling on an issue and actual, underlying support, I also collected election return data for all state ballot measures in states voting concurrently on same-sex marriage bans. Those data were obtained from the Initiative and Referendum Institute at the University of Southern California, which maintains a comprehensive database on all state ballot measures 3 .
In 30 of the 33 votes concerning same-sex marriage, the ballot measure was presented in the form of a ban on the practice. Thus, the key dependent variable used in the analyses below is the gap between preelection support for the ballot measure and actual support at the polls. For the three states that voted on legalizing same-sex marriage, the directionality of the dependent variable was adjusted accordingly to preserve consistency in the models. Polling support was calculated as the percentage of respondents supporting a ban, divided by the percentage of respondents taking a position on the matter, thus omitting undecided responses. The main hypothesis of this study is that election day opposition to same-sex marriage will be greater than that voiced in preelection polling, ceteris paribus.
Clearly, the accuracy of preelection polling may vary with a number of other factors for which we must control. For example, one would expect that surveys with larger margins of error would show greater overall discrepancies with actual election returns. Similarly, one would expect the correspondence between election returns and preelection polls to decrease as the amount of time elapsing between the poll and election day increases. Therefore, the models presented below include control variables for the margin of error for each poll as well as a variable for the number of days prior to the election that a poll was conducted. The models also include a variable for the percentage of respondents who chose “undecided” or “don’t know” in each poll, based on Berinsky’s (1999) finding that some respondents use such responses as a way of avoiding the expression of potentially controversial viewpoints.
The models also include several control variables for factors that may be related to the prevalence of social desirability bias in polling on issues of sexuality. Although the discussion above highlights the contradictory theoretical expectations we might have for some of the variables based on the current literature, it is nevertheless important to fully specify the models for state-level factors that may be reasonably related to polling bias on these issues (Egan, 2010). According to Snyder (1987), social desirability bias is greater for individuals who are highly self-monitoring, a characteristic that is more common among individuals with higher levels of education. Thus, we would expect that social desirability bias will be more prevalent in states with higher percentages of college educated citizens since those individuals are more likely to be socialized to carefully avoid expressing remarks possibly regarded as socially offensive. Therefore, the models include a measure for the percentage of a state’s residents with a BA degree or higher, obtained from the U.S. Census Bureau’s American Community Survey (U.S. Census Bureau, 2011).
The second of these control variables, the Cook Partisan Voting Index (CPVI), measures the underlying partisanship of each state’s electorate. Instead of relying on measures of self-reported party identification, the CPVI measures the average deviation of each state from the national partisan vote in the last two presidential elections. For the purposes of this study, a CPVI of zero means that a state was exactly at the national average in the last two elections in its partisan vote. A value above zero indicates a state is more Republican than the nation as a whole, while a negative value indicates it is more Democratic (Cook, 2011). The third control variable is derived from the Gallup’s annual polling on the religiosity of each of the states (Gallup, 2012b). This variable measures the percentage of a state’s residents describing themselves as “very religious.” It is expected that social desirability bias will be less common in states that are more Republican and more religious since opponents of same-sex marriage will feel less social pressure to conceal their opposition in those places. Finally, the models include a control variable for the ideological distribution of voters in each state. This variable, derived from Gallup’s polling on ideology, measures the percentage of respondents in each state describing themselves as conservative minus those labeling themselves liberal (Saad, 2009).
In addition, one of the models, described more fully below, includes a variable measuring the percentage of Americans nationwide who expressed approval for legalized same-sex marriage. This measure comes from Gallup’s annual polling on the issue (Gallup, 2012a) 4 . According to the theoretical foundations of social desirability bias and Noelle-Neumann’s “spiral of silence,” we would tentatively expect the gap between polls and election results on same-sex marriage measures would increase as more Americans support legalization of the practice, although Hopkins’s (2009) work, cited above, raises some reasons to doubt this relationship. In particular, the literature does not give us a clear expectation about whether social desirability bias is more heavily influenced by perceptions of national or localized opinion.
We must also consider the specific methodologies employed in each poll as a potential explanation for gaps between polling results and the actual vote. Egan (2010) has speculated that such gaps might be explained by flaws in the likely voter screen used by polling firms. The models specified here control for this possibility through the inclusion of a dummy variable indicating when a polling firm sampled or weighted its results based on respondents’ likelihood of voting. The other widely used technique is to sample all registered voters, which thus serves as the reference category in this study’s models. The other way this study controls for sampling method is through its primary focus on polls in which respondents were asked about other ballot measures in addition to those on same-sex marriage. Flaws with the likely voter screens would thus be expected to show distortions on nonsame-sex marriage issues as well.
Finally, Kreuter, Presser, and Tourangeau (2008) have found that in some settings social desirability bias is less prevalent in automated surveys conducted using Interactive Voice Response (IVR) technologies—also known as “robocalls”—as opposed to live interviewers 5 . Since IVR was utilized for some of the polls included in this study, the models contain a dummy variable indicating the use of this technology. Although it is expected that social desirability bias will be less common in polls using IVR, finding a statistically significant relationship would be difficult in these models since IVR was utilized in only 24 of the 136 polls examined here.
Results
In order to assess the correspondence between preelection polls and election results in same-sex ballot measures, I first examined the overall accuracy of polling on a wide range of nonsame-sex marriage ballot measures that appeared on the same ballot as same-sex marriage issues. Figure 1 displays the gaps in support between the polling and election returns for those measures by year. In general, the “error” gaps are scattered roughly evenly in a positive and negative direction. If anything, support for the general ballot measures tended to be slightly overstated in preelection polling. Overall, we see the gaps tend to cluster around zero in a relatively random fashion, as we would expect according to statistical sampling theory.
Polling-election result gap for support of nonsame-sex marriage ballot measures.
Figure 2 displays the gaps between polling and election results for only the measures regarding same-sex marriage. Even a cursory glance at this plot shows that preelection polling has almost always understated election day opposition to same-sex marriage. In fact, in only 9 of the 59 polls did election day opposition to same-sex marriage lag behind the preelection polling. This initial glance at the data suggests strong support for the hypothesis that the social desirability bias is, indeed, a factor in polling on these sorts of ballot measures.
Polling-election result gap for opposition to same-sex marriage.
In order to subject this hypothesis to more rigorous examination I estimated an OLS regression model according to the specification explained above. The unit of analysis in this model was each individual poll on any state ballot measure in a state holding a same-sex marriage vote in that election cycle. The dependent variable was the gap between polling and election day support for the “yes” vote for a particular measure 6 . The critical independent variable was a dummy variable indicating whether or not the measure dealt with same-sex marriage. Control variables, as described above, were added for each survey’s margin of error, the number of days elapsing between the poll and election day, the percentage of undecided/don’t know responses, the use of a likely voter screen, and the use of automated interviews, in addition to state-level variables for partisanship, ideology, rates of college education, and religiosity.
The regression results for Model 1, displayed in the left side columns in Table 1, are consistent with the sizable and statistically significant presence of social desirability bias in ballot measures concerning same-sex marriage. Controlling for other factors, election day opposition to same-sex marriage exceeds that expressed in preelection polling by nearly 7%. Since the dependent variable measured the election-polling gap on the “yes” vote for all ballot measures (except for the 2012 votes in Maine, Maryland, and Washington), these results are consistent with a pattern of social desirability bias and seem to mitigate against the possibility of acquiescence bias, a phenomenon is in which poll respondents are more likely to chose the agreeable, or “yes,” response (Krosnick, 1999; Ong & Weiss, 2000).
Polling-Election Results Gap for Support of State Ballot Initiatives, Regression Results.
Note: Coefficients in
In order to determine if the results were influenced by the ideological directionality of the dependent variable, I reestimated the model after recomputing the dependent variable so that positive values corresponded with the conservative ideological position on each ballot measure. As shown in the right side columns in Table 2, the regression results were almost identical to those from Model 1. The coefficient for the variable representing a same-sex marriage ballot measure was again statistically significant, indicating that opposition to same-sex marriage was over 5% greater on election day than in preelection polls. Once again, none of the control variables were statistically significant. This strengthens our confidence that the results for the same-sex marriage variable are consistent with social desirability bias rather than factors related to partisanship or ideology.
Polling-Election Result Gaps for Support of Same-Sex Marriage Bans, Regression Results.
Note: Coefficients in
As noted earlier, the theoretical expectations regarding social desirability bias and interactive effects with other contextual variables are somewhat unclear from the existing literature. For example, we might expect social desirability bias to be more prevalent in states with higher levels of college education since college-educated individuals have likely been exposed to greater socialization in the need to respect diversity and a need for more self-monitoring (Snyder, 1987). Similarly, we might expect social desirability bias to be more prevalent in polling on this issue as a greater percentage of the overall population comes to support same-sex marriage. Therefore, in order to more closely examine the theoretical underpinnings of social desirability bias as it relates to polling on same-sex ballot measures, I reestimated the analysis on the same-sex measures only. Reestimating the model for only the same-sex measures also allows us to consider several alternative explanations that might account for this pattern of polling bias.
As shown in Table 2, the results of this model disconfirm the most likely alternative explanations, specifically that the effects found above might be driven by factors related to polling methodology. For example, Egan (2010) has speculated that the discrepancy between polling and election day opposition to same-sex marriage might be attributable to problems with the likely voter screens utilized by polling firms. However, the results presented in Table 2 disconfirm that possibility since the variable representing the use of a likely voter screen is not statistically significant. Similarly, no statistically significant relationship is found between the polling-election day gap on same-sex marriage issues and a survey’s margin of error, the time elapsed since the poll was taken, nor the percentage of undecided/don’t know responses.
The results for the remaining control variables proved to be inconclusive. The only two coefficients to reach statistical significance were those for a state’s level of religiosity and Republican voting. The result for religiosity is consistent with social desirability bias. In states with larger percentages of very religious individuals, opponents of same-sex marriage will presumably feel less pressure to obscure their opinions. The results are less understandable for the variable representing the degree of Republican voting in a state, since we would likely have expected that a more Republican electorate would have the same effect as high levels of religiosity.
Finally, we turn our attention to the two variables more directly related to social desirability bias, the use of automated robocalls and the overall national level of approval for same-sex marriage. As discussed above, we would expect the use of automated robocalls to reduce the prevalence of social desirability bias in polling on this issue due to greater perceived anonymity. Similarly, we would expect social desirability bias to increase as a larger percentage of the general population comes to support same-sex marriage. In both instances, the coefficients failed to reach statistical significance. As discussed above, the use of automated interviews, although growing, is still relatively rare in opinion polling. Since robocalls were only used in 11 of the 44 polls included in the model, it is unlikely the coefficient would achieve statistical significance. Nevertheless, the coefficient was directionally consistent with an explanation of social desirability bias and came quite close to achieving statistical significance.
Conclusion
Although Hopkins (2009) has found that social desirability bias has diminished over time in mixed-race elections, the results presented above demonstrate that such effects are still very much a factor in other sorts of races, specifically those concerning same-sex marriage. Other things equal, election day opposition to same-sex marriage is between 5% and 7% greater than found in preelection polls. The results presented above are consistent with the presence of social desirability bias in polling on these issues and disconfirm the most plausible alternative explanations such as acquiescence bias and issues related to polling methodology.
It is difficult to say conclusively why social desirability bias continues to plague polling on same-sex marriage even as it has largely disappeared on issues of race and gender, but it is likely due to the fact that opinion change in the area of homosexual rights continues to lag behind those of race and gender relations. Americans, though increasingly tolerant of homosexuality, are about evenly divided on the notion of same-sex marriage. In other words, there is greater potential for social desirability bias on such issues because there is still greater division of opinion amongst the U.S. population. For example, in a Pew survey leading up to the 2008 presidential election that asked Americans whether they would consider voting for certain types of candidates, 46% of Americans said they would be less likely to vote for a homosexual candidate, whereas only 4% of Americans said they would be less likely to support a candidate who was Black (Luo, 2007). In sum, we might hypothesize a curvilinear relationship between broader measures of public opinion and social desirability bias on these sorts of sensitive issues. Such bias is likely to be less prevalent on issues with overwhelming public opposition, but increase as opponents sense their opinions are becoming socially stigmatized with a larger percentage of the public. At some point, as a new dominant consensus emerges, there are simply too few opponents remaining who might be susceptible to social desirability bias. This hypothesis could prove fruitful for further investigation.
A possible implication of these findings is that polls in recent years may be overstating levels of national support for same-sex marriage and lesbian, gay, bisexual, and transgender (LGBT) rights more generally. If social desirability bias is prevalent, as we have seen, on specific ballot measures, then it is quite possibly a problem that plagues other similar issues. For example, as noted above, Gallup found national (but narrow) majority support for the first time for legalized same-sex marriage in 2012. This study’s findings suggest that such support may be overstated by several percentage points.
Although the results presented here demonstrate a significant pattern of polling bias on issues of same-sex marriage, arriving at a convincing explanation for the specific individual-level mechanisms that lead to such biases is not yet fully complete. The findings presented above confirm that social desirability bias on these sorts of ballot measures is more prevalent in states with larger populations of Republican and highly religious voters. Nevertheless, a difficulty in studying this area is in relying on aggregate data to investigate what is essentially an individual-level phenomenon. The idea of social desirability bias postulates that survey respondents will, in some instances, express viewpoints contrary to their own in order to curry favor (or at least avoid potential disfavor) from interviewers. The specific psychological mechanisms for such behaviors are best explored through experimental research designs. What we can say is that the empirical findings at the aggregate level, examined with appropriate methodological controls, continue to be consistent with the hypothesis that social desirability effects are found in polling on ballot measures regarding same-sex marriage.
Based on prior research of Bradley effects in mixed-race elections, one might expect this phenomenon to diminish if societal attitudes continue to substantially liberalize on issues related to homosexuality. As a greater number of people come to accept same-sex marriage, there will be fewer potential polling respondents available to give misleading responses. This is a question that will be fascinating to study over time to see if it, indeed, turns out to be the case.
Footnotes
Acknowledgements
I would like to thank Nicholas Valentino, Mark Brewer, and Amy Fried, as well as the anonymous reviewers, for their many helpful suggestions during the development of this project.
Declaration of Conflicting Interests
The author declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.
Funding
The author received no financial support for the research, authorship, and/or publication of this article.