Identifying psychological responses of stigmatized groups to referendums

Significance

Voter initiatives and referendums concerning the rights of marginalized groups are utilized in 27 states, but research on the psychosocial consequences of these initiatives is underdeveloped. This research combines official records of televised advertisements from same-sex marriage campaigns with psychological outcomes among a probabilistic sample of lesbian, gay, bisexual, and transgender (LGBT) people. A novel design assesses advertisement effects on LGBT people unlikely affected by these campaigns in other ways, providing unbiased estimates. Greater ad exposure was associated with more stress among LGBT respondents; negative ads evoked sadness, whereas positive ads evoked enjoyment and happiness. No associations were observed among non-LGBT respondents. Voter referendums thus represent a source of minority stress and resilience for marginalized groups.

Abstract

Public votes and referendums on the rights of marginalized communities are utilized in 27 states and occur with some regularity. However, research has only recently begun to examine the psychological consequences of these voter referendums for members of stigmatized groups, and a number of important questions remain regarding the internal validity and generalizability of the existing evidence. The current study advances this literature by combining survey data from a large probability-based sample conducted in 2012 [lesbian, gay, bisexual, and/or transgender (LGBT) n = 939; non-LGBT n = 31,067] with media market ad-buy data in states where marriage equality was on the ballot. Television media markets cross state boundaries, ensuring that there was an unintended group of people in 12 states who were exposed to the same-sex marriage discourse but who did not live in states with the voter referendum (“media market spillovers”). We take advantage of this unique data structure by comparing LGBT people in the media market spillovers to those residing in the same state but in nonspillover markets with no ad exposure. LGBT people are emotionally affected by these campaigns, and non-LGBT people are unaffected. LGBT people in markets with a cumulative total of 400 ads have a 34.0% greater probability of reporting stress than LGBT people not exposed to ads. Additionally, while the negative ads evoked sadness, positive ads evoked enjoyment and happiness. Thus, public votes on minority rights represent both a source of minority stress and resilience.

Voter referendums have become a common way to expand or limit rights afforded to minority groups. Tracing back to 1945, Latino communities in California faced numerous initiatives and referendums denying undocumented immigrants public benefits (e.g., Proposition 187), providing English-only education programs (e.g., Proposition 227), and excluding protections in employment, housing, and public accommodations (1). By 2002, 12 states voted on 24 initiatives or referendums on abortion-related regulations, and there was a concerted effort to pass “personhood” policies through direct democratic systems to undermine the 1973 decision in Roe v. Wade (2). Direct initiatives and referendums also increased the diffusion of policy restrictions on marriages for same-sex couples (3); between 1974 and 2009, there were 158 referendums and initiatives on the rights of lesbians, gay men, bisexuals, and transgender (LGBT) people (4). Thus, voter-initiated referendums occur with some regularity and affect numerous minority groups, including people of color, women, people with a felony conviction, homeless populations, and LGBT people, among others. Despite the ubiquity of voter referendums, research regarding the psychological consequences of such referendums is underdeveloped, with most studies assessing the psychosocial consequences of voter referendums after their passage (5). The current article addresses this topic by using a unique data structure that combines official records of televised advertisements from same-sex marriage campaigns with psychological outcomes among a representative sample of LGBT people. While this study focuses on LGBT populations, it has implications for other stigmatized groups that are subject to voter-initiated referendums.

Several theories are relevant to understanding whether and how exposure to a devaluing social discourse preceding a voter referendum may affect the mental health of the stigmatized. Social identity threat theories of stigma (6) posit that cues from the social environment that are appraised as potentially harmful to one’s stigmatized social identity engender threat; this, in turn, creates involuntary stress responses and motivates a variety of coping strategies aimed at reducing identity threat. Similarly, minority stress theory (7) describes the unique, chronic, and socially based stress that individuals from stigmatized social categories are exposed to, as well as the individual- and group-level coping processes stigmatized individuals employ that may ameliorate the negative mental health consequences of these stressors. Applied to the specific context of voter campaigns, both theories suggest that campaigns seeking to undermine the rights of stigmatized individuals could potentially threaten their social identity and expose them to a variety of stressors that contribute to adverse outcomes. At the same time, these theories also underscore the existence of individuals’ coping responses to stigma-induced stressors that, if successful, could not only buffer the negative impact of these stressors but also lead to positive psychosocial outcomes (e.g., collective action). These theories thus raise the possibility that voter campaigns could generate a mix of positive and negative effects for the stigmatized.

The few studies that have examined the psychological consequences of voter campaigns for minority group members have produced findings that are consistent with the predictions of these stigma theories (6, 7). For example, Frost and Fingerhut (8) used an experience sampling design to obtain daily reports of 62 same-sex couples from four states in the month before citizens of those states voted on same-sex marriage. Even though same-sex marriage ultimately passed in these states, same-sex couples who reported more daily exposure to negative campaign messages leading up to the vote experienced increased negative affect and decreased relationship satisfaction (8), underscoring the importance of separating the effects of the voter referendum outcome from the discourse surrounding it. Another study using retrospective reports of exposure to negative campaign messages found that LGB respondents report numerous campaign-related stressors that may contribute to adverse mental health outcomes, including seeing misleading media portrayals and antigay graffiti; overhearing negative comments and jokes; experiencing hostile conversations with neighbors, colleagues, and family members; and feeling a lost sense of safety (9).

Studies have also documented positive psychological effects of voter campaigns that exist alongside these negative outcomes. In a cross-sectional survey of 354 LGB adults assessed in the 5 d leading up to the Proposition 8 referendum in California (a ballot initiative that restricted marriage to heterosexual couples), respondents reported several positive consequences of the referendum, particularly increases in LGBT community involvement and political activism, as well as feelings of pride (10). Retrospective studies have documented additional sources of resiliency associated with campaign ballot initiatives among LGB respondents, including an enhanced sense of personal and communal efficacy, the opportunity to experience personal growth (e.g., having a greater understanding of how prejudice affects their lives), and support from some heterosexual allies, which reduced the sense of isolation and powerlessness (9, 11).

These existing studies have provided important insights, but a number of critical questions remain regarding the internal validity and generalizability of the existing evidence. The findings from the present study advance this literature in both respects. Our results can be generalized to the population, as the study capitalizes on a unique opportunity in which a probability-based, nationally representative dataset was collected of both LGBT and non-LGBT people during the time when four states held voter initiatives surrounding same-sex marriage. This dataset was combined with official records of television media market ad-buy data in states where marriage equality was on the ballot. The internal validity of this study is enhanced by the fact that television media markets cross state boundaries, ensuring that there was an unintended group of people in 12 states who were exposed to the same-sex marriage discourse but who did not live in the four states with a referendum (i.e., “media market spillovers”). Among this group, we examine whether ad exposure affects three psychological outcomes—including stress, felt emotions (i.e., happiness and sadness), and behavior (i.e., laughing or smiling), and we draw comparisons to those residing in the same states but in nonspillover markets with no ad exposure. We evaluate separate effects on these outcomes for positive ads (i.e., ads affirming of same-sex marriage) and negative ads (i.e., ads devaluing same-sex marriage), which is consistent with prior studies in political science that separate ads by candidate to examine persuasion (12, 13). The coincidence of these data therefore grants a rare opportunity to examine the psychological effects of voter initiatives and referendums.

Drawing on the theories and empirical evidence reviewed above, we make the following predictions. Because voter campaigns on minority rights may represent a unique source of stress (5, 9–11, 14–16), we hypothesize that the negative discourses surrounding these campaigns will adversely affect the psychological well-being of LGBT people. Additionally, given that voter referendums can simultaneously engender some positive outcomes, such as feelings of efficacy, pride, and belongingness (9–11), we further hypothesize that positive discourses will be associated with psychological well-being among LGBT respondents. Finally, stigma theories predict (5, 6, 17), and research documents (5, 18, 19), that stigma-related stress is unique to the stigmatized, who confront adaptational demands over and above those required by similar others who are not stigmatized. For example, following constitutional amendments banning same-sex marriage, psychiatric disorders increased among LGB residents of states that enacted the bans, but not among heterosexuals living in these same states (5). Consequently, we hypothesize that our results will be specific to LGBT populations; no relationship between campaign discourses and psychological outcomes will be observed among non-LGBT respondents.

Results

The research design that we use analyzes psychological outcomes among LGBT and non-LGBT people who resided in the 12 states that share media markets with the four states that had voter initiatives on same-sex marriage (i.e., the media market spillovers) and comparing them to those that do not share those media markets. We include in the analysis all states in the continental United States except for the four states that had voter initiatives on same-sex marriage. The cumulative number of ads aired in the previous 4 wk is our independent variable, with respondents exposed to zero ads representing those in the comparison group. To account for statewide heterogeneity, state fixed effects are used; thus, the results are interpreted as within-state. Our analyses are of respondents surveyed from October 1, 2012 through November 6, 2012 (election day for the 2012 presidential election).

Exposure to Ads and Stress.

We first tested the separate effects of positive versus negative ad exposure on stress; both positively predicted stress, and the estimates for ad valence were not significantly different from each other. In addition, model assessments (i.e., Akaike and Bayesian Information Criterions) suggested that total ad exposure provided a better model fit over those that examined negative and positive ad exposure separately. Thus, we report results on the effects of total ad exposure on stress (Table 1), which, for LGBT people, was positive and significant (b = 0.51, 90% CI, 0.25, 0.77), but, for non-LGBT people, was nearly nonexistent and was not statistically significant (b = −0.001, 90% CI, −0.05, 0.05). In Fig. 1, we plotted the effect of ad exposure on the probability of reporting stress by LGBT and non-LGBT people in. LGBT people were increasingly more likely to be stressed as campaign ads increased. For instance, compared with an environment with zero advertisements, LGBT respondents living in an environment with a cumulative total of 400 ads had a 34.0% greater probability of reporting stress (90% CI, 23.4, 44.6). Non-LGBT people were −0.06% less likely to report stress with the same increase in ads, and this effect was not statistically significant (90% CI, −3.9, 3.8).

Table 1.

Logit coefficients for the effects of ads on stress

Variable	LGBT	Non-LGBT
Total ads (100s)
b	0.51 (0.16)	−0.001 (0.03)
B	0.13	−0.0
N	818	27,391
Pseudo-R2	0.11	0.04
LL	−579.96	−18,007.37

Robust SEs clustered by state and media market grouping are in parentheses. All models control for age, sex, race, and educational attainment, and state and week of interview fixed effects. Bold-faced coefficient indicates a one-tailed P < 0.05. B, fully standardized coefficients (StdXY); LL, log-likelihood.

Fig. 1.

Ad exposure and the probability of stress among LGBT and non-LGBT people; 90% confidence intervals are represented by the dashed lines.

Exposure to Negative Ads and Emotional Well-Being.

As shown in Table 2, negative ads increased the likelihood that LGBT people reported being sad (b = 1.61, 90% CI, 0.36, 2.86) and lowered the likelihood that LGBT reported smiling or laughing (b = −1.98, 90% CI, −3.35, −0.61). The effects of negative ads on feeling happy (b = 0.19, 90% CI, −2.15, 2.53) and enjoying something (b = −0.59, 90% CI, −2.63, 1.45) were not statistically significant. For non-LGBT people, there were no significant effects of negative ads on emotional well-being (Table 2). We plotted the predicted probabilities of emotional well-being as it relates to the negative ads for LGBT people in Fig. 2. Compared with an environment with zero advertisements, LGBT respondents living in an environment with a cumulative total of 200 negative ads had a 58.8% greater probability of reporting sadness (90% CI, 28.0, 89.6). Similarly, LGBT people living in an environment with a cumulative total of 200 negative ads (compared with an environment with zero) had a 68.1% lower probability of smiling or laughing (90% CI, 43.0, 93.1). The nonsignificant effects of negative ads on happiness and enjoyment among LGBT respondents were reflected in the wide confidence intervals around the predicted probabilities for these outcomes.

Table 2.

Logit coefficients for the effects of ads on emotional wellbeing outcomes

	LGBT				Non-LGBT
	Felt emotions			Behavior:	Felt emotions			Behavior:
Variable	Happy	Sad	Enjoy	smile or laugh	Happy	Sad	Enjoy	smile or laugh
Favorable ads	2.05 (0.67)	−5.99 (1.50)	3.77 (1.97)	9.12 (4.15)	0.07 (0.05)	0.002 (0.05)	−0.01 (0.05)	−0.06 (0.08)
(100s)	0.30	−0.72	0.52	0.85	0.02	0.00	−0.003	−0.02
Negative ads	0.19 (1.42)	1.61 (0.76)	−0.59 (1.24)	−1.98 (0.83)	−0.06 (0.09)	0.02 (0.05)	0.01 (0.06)	−0.03 (0.11)
(100s)	0.02	0.13	−0.05	−0.14	−0.01	0.004	0.002	−0.006
N	756	778	771	784	27,342	27,387	27,353	27,287
Pseudo-R2	0.12	0.09	0.10	0.08	0.02	0.02	0.02	0.01
LL	−317.12	−474.80	−412.71	−435.45	−9,156.8	−12,130.4	−11,320.2	−12,248.9

Robust SEs clustered by state and media market grouping are in parentheses. All models control for age, sex, race, and educational attainment, and state and week of interview fixed effects. Bold-faced unstandardized coefficients indicate a one-tailed P < 0.05. Fully standardized coefficients (StdXY) are reported below the unstandardized coefficients.

Fig. 2.

Negative ad exposure and emotional responses among LGBT people; 90% confidence intervals are represented by the dashed lines.

Exposure to Positive Ads and Emotional Well-Being.

As documented in Table 2, positive ads increased the likelihood that LGBT people reported being happy (b = 2.05, 90% CI, 0.95, 3.15), enjoying something (b = 3.77, 90% CI, 0.53, 7.01), and smiling or laughing (b = 9.12, 90% CI. 2.29, 15.95), and also decreased the likelihood that LGBT people reported being sad (b = −5.99, 90% CI, −8.46, −3.52). For non-LGBT people, there were no significant effects of positive ads on emotional well-being. We plotted the predicted probabilities of emotional well-being as it relates to the positive ads for LGBT people in Fig. 3. For each of the positive emotions, the probability of reporting emotional well-being increased for LGBT people to the point that an increase from 0 to 200 positive ads made it more likely that LGBT people reported positive emotions. Compared with an environment with zero advertisements, LGBT respondents living in an environment with a cumulative total of 200 positive ads had a 14.1% greater probability of reporting happiness (90% CI, 11.2, 16.9), a 20.9% greater probability of reporting enjoying something (90% CI, 17.5, 24.5), a 22.3% greater probability of smiling or laughing (90% CI, 19.1, 25.4), and a 27.3% lower probability of reporting sadness (90% CI, 24.2, 30.4).

Fig. 3.

Positive ad exposure and emotional responses among LGBT people; 90% confidence intervals are represented by the dashed lines.

SI Appendix, Tables S19, S21, and S22 present the results from three sensitivity analyses. These models show that ad exposure was not significantly associated with (i) an outcome that it is highly unlikely to influence (i.e., ever being diagnosed with cancer), (ii) our outcome variables among LGBT respondents who lived in matched media markets (i.e., media markets that were demographically most similar to the media markets with marriage equality ads) that were unexposed to the ads, or (iii) our outcome variables among LGBT respondents who resided in the same media markets but were assessed in an earlier timeframe (June 1, 2012 to July 31, 2012). These analyses collectively provide further evidence for the validity of our inferences.

SI Appendix also contains results from models including respondents from the four states with the voter initiatives, which show that ad exposure was not related to psychological outcomes among LGBT residents within those states (SI Appendix, Tables S4 and S5).

Discussion

Voter referendums are commonly used to expand or limit rights afforded to minority groups (1, 2, 4). Perhaps no other group is more exemplary in this regard than LGBT people. Between 1974 and 2009, there have been 158 referendums and initiatives concerning LGBT people, with over 70% of them restricting or rejecting their rights (4). It is important to examine the psychological effects of direct initiatives and referendums on marginalized groups because, even when discriminatory policies are not enacted from direct votes, the discourse from the campaign may be a unique source of minority stress and resilience (8).

Our findings indicate that total exposure to ads contributed to greater stress among LGBT respondents; further, negative ads induced a significant decrease in some positive emotions and behavior as well as an increase in negative emotions, whereas positive ads evoked greater feelings of enjoyment and happiness and decreased the likelihood of feeling sad. These results confirm prior studies documenting both positive and negative psychological outcomes—including psychological distress (8), negative affect (10), and perceived stress (14), but also feelings of pride, optimism, and communal efficacy (9–11)—among LGB individuals in response to the stigmatizing discourses that occur during voter referendums. In addition, consistent with stigma theories (6, 7, 17) and empirical evidence from the stigma literature (5, 18, 19), these psychological effects were unique to the stigmatized group targeted by the ads; no association was observed among non-LGBT respondents.

Our study improves upon the existing literature on the psychological consequences of voter referendums in four important ways. First, we relied on the largest nationally representative, probability-based sample of US adults that assessed LGBT identification, ensuring that our results have external validity. Second, our study does not ask respondents to attribute the outcomes (i.e., psychological well-being) to the exposure (i.e., marketing campaigns). Such a measurement approach, which characterizes prior work (10), introduces the possibility for confounding between the exposure and outcome. Third, we examined the direct effect of televised campaign-ad exposure to same-sex marriage initiative campaigns by linking measures of advertisement exposure via official ad-buy records by media market to survey data on psychological outcomes. This methodological approach overcomes the limitations of self-reports of campaign exposure, which introduce recall bias that can potentially overestimate effects (20). Finally, because campaigns intentionally and strategically evoke emotional responses from voters (21), we focused our analysis on a subpopulation of people who were unintentionally exposed to campaign ads, thereby reducing selection bias.

Our failure to find psychological effects of ads among LGBT residents intentionally exposed to campaign ads (i.e., those living within the four states that voted on marriage equality) may be due to the small sample size of LGBT respondents in these states, which substantially reduced our statistical power. It is also possible that living in states that voted on marriage equality afforded LGBT individuals some agency in affecting the outcome. Support for this possibility comes from research showing that LGB respondents living in states that passed amendments banning same-sex marriage reported significantly more days engaged in LGBT rights activities and were nearly 3 times more likely to vote than LGB respondents in states without the amendments on the ballot (15). Research with other minority groups, including Latinos (22) and Asian Americans (23), has similarly documented that ballot initiatives targeting these groups led to their greater political involvement. The positive effects of such collective action in counteracting stigma (24) may have neutralized other negative psychological consequences of the media campaigns for the LGBT respondents in these four states.

While advancing the literature, the present study has some limitations. The reliability of the dependent variables has not been evaluated; however, they correlate in expected ways with a measure that has undergone such examination (SI Appendix), providing some preliminary support for construct validity. Nevertheless, future research should replicate these findings with more comprehensive measures of stress and mental health. In addition, we are unable, with these data, to unpack the mechanisms through which ad exposure affects the psychological outcomes, although prior studies have suggested some psychosocial processes through which voter referendum may affect the mental health of LGB respondents (9, 10). It is also unknown from the current design how durable these effects might be; future longitudinal research is needed to address this question. Finally, our research sought to address the overall impact of voter campaigns on the mental health of LGB respondents. Consequently, we operationalized ad exposure as aggregate summaries of media market buys. While this approach uses an official tabulation of ad exposure, it does not attempt to identify the effects of specific ads or discourses contributing to study outcomes, nor does it document actual observation of the ads.

Voter initiatives and referendums exist in 27 states, and the ballot box is regularly used to confer or restrict rights to marginalized members of society. While we focused on LGBT individuals, other groups have recently been targeted by voter campaigns, including immigrant populations (e.g., Measure 58 in Oregon on English-only education) and Muslims (e.g., State Question 755 in Oklahoma on Islamic law in state courthouses), among others. Research points to several dimensions along which stigmatized groups may differ, such as stereotypes regarding warmth and competence (25) as well as perceptions regarding the concealability and disruptiveness of the stigma (26). These factors are likely to influence the content and scope of voter campaigns targeting stigmatized groups. Whether the distinct content and scope of different voter campaigns alters their psychological consequences for members of stigmatized groups represents an important direction for future research on the generalizability and potential boundary characteristics of the results reported here.

Materials and Methods

Design, Sample, and Survey Data Collection.

Respondents come from a nationally representative sample, the Gallup Daily tracking poll, which measures political and social attitudes, well-being, and demographic characteristics of the US adult population (27). The poll interviews, on average, 1,000 American adults every day, excluding major national holidays. Although the daily response rates average 11% (28), previous studies show that low response rates minimally affect survey accuracy, especially when analyzing lifestyle, health, and demographics (29). Computer-assisted landline telephone and cell phone interviews are conducted in both English and Spanish, with adult respondents randomly selected within contacted households.

The survey interviewed a total of 373,352 adults from June 1, 2012 through June 30, 2014. We constrain our analysis from October 1, 2012 to November 6, 2012 (election day) because this is the period of time when political campaigns are most active; in this timeframe, a total of 32,006 adults were interviewed in the continental United States. Respondents are from all 50 states and the District of Columbia. The weighted and adjusted sample represents ∼98% of the US adult population (27). The data are fully deidentified, and, since the analysis of these data are secondary, this study did not require review by the institutional review board.

Measures.

Dependent variables.

Five questions from the Emotional Health Index subset of the Gallup-Healthways Well-Being Index assessed psychological outcomes and stress responses. Respondents were told, “Please think about yesterday, from the morning until the end of the day. Think about where you were, what you were doing, who you were with, and how you felt.” Afterward, respondents were asked, “Did you smile or laugh a lot yesterday? Did you experience the following feelings a lot of the day yesterday? How about sadness? How about stress? How about happiness?” We dichotomize responses as 0 (no) and 1 (yes) for stress, emotions (sadness, happiness), and behaviors (smiling or laughing). SI Appendix, Table S19 shows the correlation between our outcome measures and a well-validated measure of health (self-rated health), providing some evidence for the validity of these measures.

Sexual orientation.

In June 2012, the Gallup Daily tracking survey started to ask the question, “Do you, personally, identify as lesbian, gay, bisexual, or transgender?” Between 2012 and 2014, 3.8% of adults in the survey identified as LGBT, which is similar to and within the 95% confidence interval of the 3.7% (2.7 to 4.7%) estimate from National Opinion Research Center’s 2012 General Social Survey (GSS) (author calculation). While the 2012 GSS has a smaller total sample size (n = 1,974) than the Gallup data, it has a much higher response rate at 71.4% (30). The demographic composition of LGBT respondents regarding age, gender, and race/ethnicity is similar across several population-based surveys, including Gallup’s data (31).

We constrain our analyses to respondents who identified as LGBT and those who did not; consequently, respondents who refused to answer the question or reported “Don’t know” are removed. We did not conduct any imputation of the LGBT identity variable for several reasons. First, studies suggest that most nonresponders to sexual orientation questions would most likely not be classified as LGBT (32). Second, imputation of the LGBT sample introduces an uncertainty that seems unwarranted given sufficient sample size for these analyses. For the present study, LGBT individuals are those who answered yes to this question (n = 939), and non-LGBT individuals are those who answered no (n = 31,067).

Independent variable.

Our independent variable is official records of televised ads from same-sex marriage campaigns in 2012. Political campaigns purchase ads targeting designated media markets (DMAs) within specific states. DMAs are geographic boundaries that define a specific television media market. Each DMA has unique programming and advertisements, and they are determined by the Nielsen Group. There are 210 DMAs that cover the contiguous United States, Hawaii, and portions of Alaska. A grouping of counties forms most DMAs, so we nest respondents, based on their county of residence, into their respective DMA. While DMAs are determined by counties, they are seldom determined by statewide boundaries, and there are often DMAs that cross state lines.

A record of political ad buys is collected by the Campaign Media Analysis Group (CMAG) Advertising Data Report and Ad Alerts. CMAG classifies campaign ads as they are broadcast, records each time they appear by DMA, and maintains documentation relating to organizational sponsors of the ads. We acquired the ad buys relating to marriage equality initiative campaigns in the 2012 presidential election from CMAG. There were 14 DMAs that had ads aired that targeted Maine, Maryland, Minnesota, and Washington, the four states with referendums on same-sex marriage. These 14 DMAs crossed over into 12 additional states that did not have a direct initiative relating to same-sex marriage, providing the market spillovers. SI Appendix, Table S7 provides the summary statistics of ad exposure by state of residence and LGBT status; SI Appendix, Fig. S8 depicts the distribution of ad exposure in the media market spillovers.

The CMAG dataset records the sponsor of the ad (e.g., “Mainers United for Marriage Equality” and “Protect Marriage Maine”). Ads are classified as positive if they are favorable to marriage equality and negative if they are opposed to marriage equality. We determined that each ad’s sponsor would indicate whether the ad was positive or negative (i.e., campaigns in favor of marriage equality would produce ads supportive of marriage equality and vice versa). The content of the ads in favor of marriage equality embraced values of love and fairness, and the content of the ads opposed to marriage equality emphasized the loss of liberties people of faith and parents would face if same-sex marriages were introduced. Thus, the ads in favor of same-sex marriage, from the perspective of LGBT rights advocates, incorporated empowering discourses, while ads in opposition used devaluing discourses.

There are two ways previous scholarship has operationalized exposure to televised advertising over a specified window of time: running totals of gross ratings points (GRPs) and running totals of the numbers of ads aired (12, 13). (GRPs measure the size of the audience likely exposed to advertisements.) A previous study (13) showed that there is a strong correlation between these two measures ($r=0.95$). The most often selected timeframe is a 4-wk timespan to measure ad volume, and a previous study tested different timeframes for sensitivity, finding 4 wk to be ideal (12). CMAG does not provide GRPs without the research team having an additional subscription to the Nielsen Group, so they are not available for this study. Thus, we use the running total of the number of positive and negative ads aired in the previous 4 wk. Since these two measures are highly correlated and have both been used in previous research, we do not anticipate running totals to mischaracterize ad exposure. In the 2012 election, the pro-marriage equality side delivered substantially more ads (n = 12,756) than the anti-marriage equality side (n = 9,525). The vast majority of these ads aired in the month before election day; since we restricted our analyses to October 1 through election day, our use of a 4-wk running total includes all ads aired from September to November. The cumulative number of ads aired in the previous 4 wk in a respondent’s media market is calculated based on the day each respondent was interviewed. There were 511 ads aired in the month of August in Minnesota media markets that are removed from analysis. (All of the Minnesota ads were in favor of marriage equality and aired starting August 16, 2012 in the Duluth and Minneapolis media markets.)

We focus on two explanatory variables relying on the CMAG ad buys: (i) a running total of the favorable campaign ads aired and (ii) a running total of the negative campaign ads aired. For the analyses related to stress, we combined these two variables and reported total ad volume, because both variables positively predicted stress and the estimates for ad valence were not significantly different from each other.

Analytical Strategy.

Our research design enables us to identify an effect of direct initiatives and voter referenda on psychological outcomes. Television media markets are unique in that they cross state boundaries. For example, residents of Benewah County, Idaho, or Wallowa County, Oregon, are in the Spokane Media Market, which is one of the four markets servicing Washington State. These overlapping state boundaries mean that people may reside in media markets airing campaign ads but who were unintentionally exposed because they did not reside in a state with a referendum on the ballot. In this example, residents of Benewah County, Idaho, were exposed to ad campaigns related to the same-sex marriage initiative but were living in a state that was not voting on the initiative. Analyzing the psychological outcomes among this subset of individuals—what we call “market spillovers”—therefore removes multiple sources of unobserved confounding that are associated with both our independent variable (ad exposure) and dependent variables (psychological outcomes and stress). These confounders include other campaign-related materials (e.g., leaflets, canvassers, and phone calls), receiving empowering or devaluing information about LGBT people in the mail, or having strangers attempt persuasive conversations about marriage equality. Additionally, televised campaign ads are designed to evoke emotional responses (21) and to microtarget persuadable groups of registered voters (33). Markets may have been selected to maximize electoral gains by evoking emotions, which means that the dependent variables may have been selected on determining how many ads to air and which particular ads to air to evoke psychological responses.

We remove from the analysis any respondent who resides within one of the four states voting on marriage equality (Maine, Maryland, Minnesota, and Washington). This way, we remove all of these other marriage equality campaign-related activities (e.g., canvassing, phone banking, and mailed leaflets), which were not present in states without the voter referendums on the ballot. This removal also addresses potential selection bias. Certain DMAs within states with voter referendums may have been targets for campaign ads based on the size of the market relative to the state population and also strategic investment into markets that may be more responsive to campaign appeals. Individuals residing in the DMA but outside the state that is voting on the referendum are accidentally targeted with campaign ad material. The Gallup data are beneficial because a substantial portion of the survey sample resides in DMAs outside of the 14 DMAs where there were marriage equality campaign ads. This leaves a sizeable portion of the sample as a comparison group unexposed to ads. An advantage of using campaign-ad data is that we do not rely on self-reported exposure to televised ads, which has the tendency to overestimate effects (20). We cannot, however, verify whether individuals in the spillover DMAs actually viewed the televised ads; thus, the effects we report are equivalent to intent-to-treat effects.

We employ state fixed effects and weekly fixed effects. The statewide fixed effects control for state attributes that may be a source of heterogeneity among respondents, and this makes inferences within-state. The weekly fixed effects account for other potential campaign events or events that received national attention during the month leading up to election day, such as the reelection of Barack Obama. All of the SEs are robust and clustered by state–DMA grouping to account for nonindependence at that level. This analytical strategy is consistent with previous studies that exploit the accidental treatment of campaign ads that spill over in neighboring states (12).

The outcome measures are coded dichotomously; thus, we use logistic regression models throughout. Each model controls for demographics including age (continuous), whether respondents have obtained at least a college degree (yes/no), sex (male/female), and race or ethnicity (non-Hispanic black, Hispanic, and Other). The survey weights are used in all of the analyses presented here. The survey weights account for unequal probabilities of selection and survey nonresponse, and they are poststratified for demographic groups to match that of population estimates. We conduct analyses separately for LGBT and non-LGBT adults. Sample sizes differ across the models for one of two reasons: instances of perfect prediction were removed from the analysis, and item nonresponse on the dependent variable. After estimation of the logistic regression models, we report predicted probabilities by fixing ad volume at specific levels, predicting probabilities of outcomes for respondents as they are observed in the dataset, and then averaging these predicted probabilities (i.e., average marginal effects at representative values). We report 90% confidence intervals for our point estimates and predicted probabilities to approximate a one-tailed P value of 0.05, given our directional hypotheses. [All of the results except one would be considered statistically significant with two-tailed tests (see SI Appendix, Tables S1–S4).] All analyses are conducted in Stata v. 14.

Additional model output and predictions are provided in SI Appendix, including summary statistics, further details on ad exposure, assessments of the dependent variables, sensitivity analyses, and model results from ordinary least squares regressions.

Both the Gallup Daily tracking survey and the CMAG datasets are proprietary, and researchers have to agree to data use and sharing agreements. The data used in this study cannot be made publicly available due to these copyright restrictions. However, the command scripts are provided in SI Appendix, which permits replication of our findings and summary statistics. Researchers interested in replicating the results should consult the corresponding author to purchase the datasets from the Gallup Organization and CMAG.