Peer Victimization and Sexual Risk Differences Between Lesbian, Gay, Bisexual, Transgender, or Questioning and Nontransgender Heterosexual Youths in Grades 7–12

Abstract

Objectives. Before and after accounting for peer victimization, we estimated sexual risk disparities between students who self-identified as lesbian, gay, bisexual, transgender, or questioning (LGBTQ) and students who self-identified as nontransgender heterosexual.

Methods. Students in grades 7 through 12 in Dane County, Wisconsin, were given the Web-administered Dane County Youth Assessment. One set of analyses was based on a sample that included 11 337 students. Subsequent analyses were based on a sample from which we screened out students who may not have been responding to survey items truthfully. Various multilevel-modeling and propensity-score-matching strategies ensured robustness of the results, examined disparities at lower and higher victimization rates, and explored heterogeneity among LGBTQ-identified youths. Finally, propensity-score-matching strategies estimated LGBTQ–heterosexual disparities in 2 matched samples: a sample that reported higher victimization and one that reported lower victimization.

Results. Across 7 sexual risk outcomes, and in middle and high school, LGBTQ-identified youths reported engaging in riskier behavior than did heterosexual-identified youths after we accounted for peer victimization. Risk differentials were present in middle and high school. The LGBTQ group was heterogeneous, with lesbian/gay- and bisexual-identified youths generally appearing most risky, and questioning-identified youths least risky. In the matched sample with lower average victimization rates, LGBTQ-identified youths perceived a greater risk of sexually transmitted infections despite not engaging in sexually risky behavior at significantly higher rates; in the matched sample with higher average victimization rates, all outcomes were significantly different.

Conclusions. Demonstrated LGBTQ–heterosexual risk differentials in grades 7 through 8 suggest that interventions need to be implemented during middle school. These interventions should also be differentiated to address the unique risk patterns among LGBTQ subgroups. Finally, models of sexual risk disparities must expand beyond peer victimization.

High school–aged youths who self-identify as lesbian, gay, bisexual, transgender, or questioning (LGBTQ) tend to exhibit more risk with regard to sexual behavior (such as the number of sex partners, use of protection, and prevalence of sexually transmitted infections [STIs]) than do their self-identified, nontransgender heterosexual peers.^1,2 Across middle and high school, LGBTQ youths are also disproportionately the victims of bullying.^3–7 Given LGBTQ youths’ high levels of both sexual risk and victimization, a recent meta-analysis suggested that abuse (including peer victimization, but also parental physical abuse and childhood sexual abuse) may actually contribute to sexual risk disparities between LGBTQ and heterosexual youths.²

Victimization may lead to increased sexually risky behavior through several pathways, including heightened feelings of isolation leading to increased desires for sexual fulfillment,⁸ psychological distress leading to unprotected sex,⁹ and stigmatization leading to lack of knowledge about sexual-minority identity¹⁰ in turn leading to seeking out same-sex activities to navigate identity. In a recent retrospective study of LGBT young adults, those who recalled greater peer victimization during their teenage years were more likely to have ever had an STI and to have been at HIV risk in the past 6 months.¹¹ The focus on victimization has also caught on in the media and in some education reform.^12,13 However, researchers have noted that victimization may be only one component in explaining LGBTQ–heterosexual suicide-related risk disparities^14,15; thus, it stands to reason that victimization may explain part of LGBTQ–heterosexual sexual risk disparities, but that substantial disparities may remain after victimization is accounted for.

We sought to provide a rigorous assessment of sexually risky attitudes and behaviors before and after accounting for peer victimization among a large population-based sample of LGBTQ- and heterosexual-identified youths in middle and high school. In the process, we implemented a novel procedure for screening out potentially mischievous survey responders and a series of propensity-score-matching analyses to estimate LGBTQ–heterosexual sexual risk disparities among matched LGBTQ- and heterosexual-identified students with the same demographic characteristics and levels of peer victimization. The methodological advances implemented here provide a more nuanced view of LGBTQ-identified youths and, thus, may improve the effectiveness of future prevention efforts.

METHODS

Data were from the Dane County Youth Assessment (DCYA), an anonymous Web-based survey (2008–2009) administered in schools to students in grades 7 through 12 in Dane County, Wisconsin. The final analytic data set contained a total of 11 337 students (n = 3031 in middle school [grades 7–8]; n = 8306 in high school [grades 9–12]) in 30 schools. The DCYA is a countywide collaborative project among schools and several community organizations (e.g., United Way, Department of Human Services). The county represents geographically diverse areas ranging from small working farms to a large city. The percentage of students receiving free or reduced-price lunch ranged from 16% to 58% across the schools. The survey assessed a wide range of physical and mental health indicators, as well as various attitudes and social behaviors. Students completed these anonymous surveys independently while in school during proctored sessions. A waiver of active consent was employed, and child written assent was used. Surveys were given to all students in grades 7 through 12; the response rate was very high, ranging from 90% to 95% across the 30 schools.

We first performed analyses on the full sample (model 1, described in the “Analytic Plan” section), and then we restricted the sample to exclude students who provided unusual response patterns (i.e., multiple low-frequency responses to questions that are in principle unrelated to sexual orientation and transgender identity; see Tables A and B, available as a supplement to the online version of this article at http://www.ajph.org). This more-restricted sample consisted of 11 033 students, and provided a more conservative estimate of LGBTQ–heterosexual risk differentials.⁴ Overall, 2.7% of responses (304 of 11 337) were screened out by the multiple-low-frequency responder exclusion criterion. As we expected, disproportionately more youths who ostensibly identified as LGBTQ were screened out (9.9%; 58 of 588), compared with youths who ostensibly identified as heterosexual (2.3%; 246 of 10 749). Additional details on the data, screener items, and response patterns to the screener items are contained in the supplemental document (available as a supplement to the online version of this article at http://www.ajph.org).

LGBTQ Identification

In the demographic section of the survey, students were asked to identify their sexual orientation or gender identity as gay, lesbian, bisexual, questioning, transgender, or “none of the above.” Students could choose more than 1 category. As shown in Table D (available as a supplement to the online version of this article at http://www.ajph.org), 530 students (4.8% of the analytic sample for models 2–4) selected at least 1 LGBTQ category. We considered students who selected “none of the above” and did not also select an LGBTQ category to be nontransgender “heterosexual-identified” (n = 10 503).

Victimization

The University of Illinois Victimization Scale¹⁶ was included in the DCYA. This 4-item scale assesses victimization from peers. Students were asked how often the following happened to them in the past 30 days: “Other students called me names”; “Other students made fun of me”; “Other students picked on me”; and “I got hit and pushed by other students.” Response options were never, 1 or 2 times, 3 or 4 times, 5 or 6 times, and 7 or more times. The construct validity of this scale has been supported by exploratory and confirmatory factor analysis.¹⁶ The Cronbach α for this sample was 0.86.

A second set of items asked students how often they were bullied in the past 12 months (1) “through the Internet or text messaging,” (2) “about being perceived as gay, lesbian, or bisexual,” (3) “about your race or ethnic background,” and (4) “about how you look.” Response options were never, rarely, sometimes, often, and very often. We did not collapse these items in a single index because the question on the Internet and text messaging pertained to a mode of bullying, not a type of identity- or looks-based bullying; moreover, the identity- or looks-based forms of bullying conceivably pertained to different groups.

Finally, several items in the DCYA gauged a broader set of experiences that may be considered as victimization or abuse.^17–19 These 4 items concerned (1) whether the student was kicked out of his or her house by a parent or guardian, (2) physical abuse by a parent, (3) sexual abuse by an adult, and (4) dating violence. We entered these items separately in our analysis (i.e., they were not collapsed into a single index) because they reference different types of victimization that do not necessarily represent a single construct. We also included a measure of the student’s perception of his or her parents’ love and support. Exact phrasing and response options for each item mentioned here are contained in the supplemental document.

Outcomes

We examined LGBTQ–heterosexual differences in 8 outcomes:

1. engaging in sexual activities,

2. having any sex partners,

3. having any anonymous sex partners,

4. having sex under the influence of drugs or alcohol,

5. not using protection (among the subset of students who have had sex),

6. having been tested for an STI,

7. perception of the student’s own likelihood for getting an STI, and

8. knowledge of condom effectiveness (Figures 1 and 2).

FIGURE 1—

LGBTQ–heterosexual odds ratios (OR), by model, for outcomes (a) sexual activity, (b) sex partners, (c) unprotected sex, and (d) sexually transmitted infection testing: Wisconsin, Dane County Youth Assessment, 2008–2009.

Note. LGBTQ = lesbian, gay, bisexual, transgender, or questioning. Whiskers indicate 95% confidence intervals (CIs). Models 1 through 4 and 7 were estimated via the XTLOGIT command in Stata (StataCorp LP, College Station, TX) with school random effects and use of school-cluster jackknifed standard errors. Models 5 and 6 were estimated via hierarchical linear model software and use school-cluster robust standard errors. Model 1 presents unadjusted odds ratios, including students who provided multiple-low-frequency responses on screener items. Model 2 is similar to model 1, except the multiple-low-frequency responders are removed. Model 3 conditions on race, gender, grade, age, height, weight, 2 proxies of socioeconomic status, disability status, and whether students provided 1 vs 0 low-frequency responses. Model 4 adds linear and quadric terms for a victimization composite scale, and a series of indicator variables for all possible responses to frequency of bullying via text messaging or Internet, and bullying on the basis of perceived sexual orientation, race/ethnicity, and looks. Model 5 is estimated on a propensity-score-matched sample with exact matching within schools, matched at the LGBTQ values; once matched and with balance achieved, the estimation model is a multilevel model, conditioning on all covariates included in the matching stage,²⁰ and allowing for between-school variation in the intercept, LGBTQ slope, and propensity-score slope, and clustering standard errors at the school level. Model 6 is similar to model 5, with matching done at the heterosexual values instead of the LGBTQ values, and matches from other schools within the county were allowed. Model 7 is a covariate-adjusted multilevel model, building off of model 4 and adding covariates for perceptions of family support, being kicked out of home, physical abuse by parents, sexual abuse by adults, and dating violence. For presentation purposes, 95% CIs with values above 5 were capped at 5.

FIGURE 2—

LGBTQ–heterosexual odds ratios (OR), by model, for outcomes (a) anonymous sex partners, (b) sex under the influence, (c) perceived likelihood of sexually transmitted infection, and (d) incorrect condom knowledge: Wisconsin, Dane County Youth Assessment, 2008–2009.

Note. LGBTQ = lesbian, gay, bisexual, transgender, or questioning. Whiskers indicate 95% confidence intervals (CIs). Models 1 through 4 and 7 were estimated via the XTLOGIT command in Stata (StataCorp LP, College Station, TX) with school random effects and use of school-cluster jackknifed standard errors. Models 5 and 6 were estimated via hierarchical linear model software and use school-cluster robust standard errors. Model 1 presents unadjusted odds ratios, including students who provided multiple-low-frequency responses on screener items. Model 2 is similar to model 1, except the multiple-low-frequency responders are removed. Model 3 conditions on race, gender, grade, age, height, weight, 2 proxies of socioeconomic status, disability status, and whether students provided 1 vs 0 low-frequency responses. Model 4 adds linear and quadric terms for a victimization composite scale, and a series of indicator variables for all possible responses to frequency of bullying via text messaging or Internet, and bullying on the basis of perceived sexual orientation, race/ethnicity, and looks. Model 5 is estimated on a propensity-score-matched sample with exact matching within schools, matched at the LGBTQ values; once matched and with balance achieved, the estimation model is a multilevel model, conditioning on all covariates included in the matching stage,²⁰ and allowing for between-school variation in the intercept, LGBTQ slope, and propensity-score slope, and clustering standard errors at the school level. Model 6 is similar to model 5, with matching done at the heterosexual values instead of the LGBTQ values, and matches from other schools within the county were allowed. Model 7 is a covariate-adjusted multilevel model, building off of model 4 and adding covariates for perceptions of family support, being kicked out of home, physical abuse by parents, sexual abuse by adults, and dating violence. For presentation purposes, 95% CIs with values above 5 were capped at 5.

To facilitate interpretation and estimation, we dichotomized the outcome variables into “no reported risk” and “some reported risk.” For example, concerning number of sex partners, responses of “none” would be coded as “no reported risk,” whereas responses of “1 to 2 partners” or more would be coded as “some reported risk.”

Analytic Plan

For each dichotomous outcome, we estimated a series of models that examined LGBTQ–heterosexual odds ratios (ORs) and average marginal effects (AMEs); each model was a hierarchical linear model with a logit link function, accounting for students nested within schools and using jackknifed school-clustered standard errors.²¹ We constructed the ORs and AMEs from these hierarchical linear model estimates. We first present the unconditional odds ratios and AMEs in the full analytic sample (model 1), then in the analytic sample that excludes multiple-low-frequency responders identified by the screener tool developed by Robinson and Espelage⁴ (model 2) to reduce the effects that potentially mischievous responders have on the estimates of LGBTQ–heterosexual disparities. Proceeding with the sample excluding multiple-low-frequency responders, we conditioned on demographic factors (model 3) and then added covariates for peer-based victimization (model 4).

To avoid making strong functional form and extrapolation assumptions, in a separate set of analyses we implemented propensity-score-matching approaches, to match LGBTQ- and heterosexual-identified students with similar demographic and victimization profiles.²² We generated propensity scores using the set of covariates included in model 4. We then constructed 2 samples: (1) LGBTQ- and heterosexual-identified youths matched at the LGBTQ values (i.e., finding heterosexual-identified youths who had demographic profiles and victimization levels nearly identical to those of the LGBTQ-identified youths [n = 1511; model 5]) and (2) LGBTQ- and heterosexual-identified youths matched at the heterosexual values (i.e., finding LGBTQ-identified youths who had demographic profiles and victimization levels nearly identical to those of the heterosexual-identified youths [n = 10 964; model 6]). After matching, we achieved balance and reduced the absolute standardized bias substantially for the collection of covariates used for matching (Figure A and Table E, available as a supplement to the online version of this article at http://www.ajph.org). We then used the hierarchical linear model to estimate odds ratios on each matched sample; model 5 gives the estimated odds ratios among heterosexual- and LGBTQ-identified youths who on average fit the LGBTQ-identified profile (e.g., higher average victimization levels), and model 6 gives the estimated odds ratios when students on average fit the heterosexual-identified profile (e.g., lower average victimization levels). See Table E for the characteristics of each matched sample.

The focus of our analysis was the extent to which peer victimization may explain LGBTQ–heterosexual risk disparities. However, other forms of victimization likely contributed to these risk disparities^2,17,23; thus, we present a supplemental analysis to assess whether a more comprehensive set of victimization variables, as well as a measure of perceived parental support, explains part of the remaining LGBTQ–heterosexual risk disparities. Model 7 is a covariate-adjusted multilevel model that builds on model 4 by including covariates for sexual abuse by adults, perceptions of love and acceptance by one’s parents, an indicator for whether the student was ever kicked out of the home by a parent or guardian, and an indicator for being a victim of dating violence.

Note that our sample included both middle school– and high school–aged youths, whereas previous studies have focused on high school youths^1,23 or have not disaggregated results by middle and high school.²⁴ We reestimated models 1 through 4 and 7, this time interacting middle school with LGBTQ and estimating the odds ratios separately for middle school and high school.

Finally, researchers recognize the need for studying heterogeneity within the broad LGBTQ category, particularly noticing that bisexual-identified youths tend to have elevated risk compared with lesbian/gay-identified youths.^4,25,26 Note too that the DCYA asked about whether students in this population-based sample identified as questioning their sexual orientation, and thus the sample is more likely to reflect the continuum of sexual identification than would a targeted sample of identified LGBT (not questioning) youths. We present the ORs and AMEs from models 1 through 4 and 7 for each LGBTQ subgroup: lesbian or gay, bisexual, transgender, and questioning.

RESULTS

Model 1 in Figures 1 and 2 shows that the LGBTQ-identified youths had riskier outcomes than did heterosexual-identified youths before we screened out potentially mischievous responders (i.e., multiple-low-frequency responders), statistical conditioning, or propensity score matching. For example, the odds that an LGBTQ-identified youth engaged in sexual activity was 2.62 times the odds for a heterosexual-identified youth (F_1,29 = 59.49; P < .001; 95% confidence interval [CI] = 2.03, 3.37). The corresponding AME of 0.16 for model 1 in Table 1 indicates that on average the likelihood of engaging in sexual activity was 16 percentage points higher for LGBTQ-identified youths than for heterosexual-identified youths. (The patterns and inferences in the ORs and AMEs were similar. We focus on interpreting the ORs in the rest of this article; however, readers interested in AMEs can find them in Table 1.)

TABLE 1—

Average Marginal Effects (AMEs), by Outcome, LGBTQ Subgroup, and Model: Wisconsin, Dane County Youth Assessment, 2008–2009

Outcome	Model 1, AME	Model 2, AME	Model 3, AME	Model 4, AME	Model 7, AME
Sexual activity
LGBTQ	0.16*	0.13*	0.15*	0.13*	0.08*
Lesbian/gay	0.22*	0.18*	0.16*	0.16*	0.14*
Bisexual	0.23*	0.20*	0.20*	0.18*	0.10*
Transgender	0.13*	0.04	0.05	0.06	0.03
Questioning	0.06	0.03	0.07*	0.06	0.04
No. of partners (any)
LGBTQ	0.13*	0.11*	0.13*	0.12*	0.08*
Lesbian/gay	0.20*	0.16*	0.16*	0.17*	0.15*
Bisexual	0.18*	0.15*	0.17*	0.15*	0.09*
Transgender	0.15*	0.07	0.10	0.10	0.08
Questioning	0.04	0.01	0.04	0.03	0.02
No. of anonymous sex partners (any)
LGBTQ	0.06*	0.04*	0.06*	0.05*	0.03*
Lesbian/gay	0.08*	0.06*	0.07*	0.07*	0.05*
Bisexual	0.07*	0.05*	0.08*	0.07*	0.04*
Transgender	0.08*	0.03	0.04	0.04	0.02
Questioning	0.05*	0.03*	0.05*	0.04*	0.03
Sex under the influence of drugs or alcohol
LGBTQ	0.06*	0.05*	0.08*	0.07*	0.04*
Lesbian/gay	0.08*	0.05*	0.07	0.08	0.05
Bisexual	0.08*	0.07*	0.11*	0.10*	0.06*
Transgender	0.08*	0.04	0.06	0.06	0.04
Questioning	0.03	0.01	0.04*	0.03*	0.02
Unprotected sex (among those who had sex)
LGBTQ	0.16*	0.16*	0.13*	0.13*	0.11*
Lesbian/gay	0.34*	0.36*	0.33*	0.34*	0.30*
Bisexual	0.15*	0.16*	0.12*	0.13*	0.10
Transgender	0.09	−0.09	−0.05	−0.06	−0.08
Questioning	0.19*	0.14*	0.11	0.12*	0.11*
STI testing
LGBTQ	0.05*	0.04*	0.05*	0.04*	0.02
Lesbian/gay	0.06*	0.05*	0.05*	0.05*	0.04
Bisexual	0.07*	0.06*	0.07*	0.06*	0.03*
Transgender	0.04*	0.02	0.04	0.04	0.03
Questioning	0.00	−0.01	−0.01	−0.01	−0.02
Perceived likelihood of STI
LGBTQ	0.17*	0.14*	0.12*	0.11*	0.07*
Lesbian/gay	0.20*	0.15*	0.13*	0.12*	0.09
Bisexual	0.16*	0.14*	0.11*	0.11*	0.08*
Transgender	0.21*	0.08	0.08	0.06	0.04
Questioning	0.18*	0.13*	0.11*	0.09*	0.04
Incorrect condom knowledge
LGBTQ	−0.08*	−0.08*	−0.08*	−0.08*	−0.08*
Lesbian/gay	−0.07	−0.10	−0.06	−0.07	−0.06
Bisexual	−0.11*	−0.12*	−0.10*	−0.10*	−0.09*
Transgender	−0.05	−0.01	−0.03	−0.04	−0.05
Questioning	−0.03	−0.04	−0.07	−0.07	−0.08*

Note. LGBTQ = lesbian, gay, bisexual, transgender, or questioning; STI = sexually transmitted infection. Models 1 through 4 and 7 were estimated via the XTLOGIT command in Stata (StataCorp LP, College Station, TX) with school random effects and use of school-cluster jackknifed standard errors. Model 1 presents unadjusted average marginal effects, including students who provided multiple-low-frequency responses on screener items. Model 2 is similar to model 1, except the multiple-low-frequency responders are removed. Model 3 conditions on race, gender, grade, age, height, weight, 2 proxies of socioeconomic status, disability status, and whether students provided 1 vs 0 low-frequency responses. Model 7 is a covariate-adjusted multilevel model, building off of model 4 and adding covariates for perceptions of family support, being kicked out of home, physical abuse by parents, sexual abuse by adults, and dating violence.

*P < .05.

Excluding students who provided multiple low-frequency responses to screener items, model 2 found that LGBTQ–heterosexual disparities reduced slightly, but remained highly significant across all outcomes (each P < .001), with ORs between 1.88 and 3.08 (except for “incorrect condom knowledge,” which we discuss later in this section). Note that we ran all subsequent models on the sample that excluded the multiple-low-frequency responders. Conditioning on a host of demographic covariates (in model 3) had relatively little effect on the ORs for sexual activity, sex partners, or sex under the influence.

Adding covariates for bullying victimization (in model 4) tended to account for an additional portion of the LGBTQ–heterosexual risk disparities (except for the case of unprotected sex); however, the ORs ranged between 1.70 and 2.52, and they all remained statistically significant (each P < .003). We found similar patterns when we performed analyses using both y-standardization approaches and linear probability models.^27,28 In separate analyses, we found no evidence that LGBTQ–heterosexual disparities in model 4 varied significantly by sex, with the exception of “engaging in sexual activities”: although significant among both boys and girls, the LGBTQ–heterosexual disparity was more pronounced (P = .039) among girls (OR = 2.67; P < .001; 95% CI = 1.96, 3.64) than among boys (OR = 1.59; P = .025; 95% CI = 1.06, 2.37).

Models 5 and 6 present the ORs estimated on the propensity-score-matched samples. When we compared LGBTQ-identified youths with their matched heterosexual-identified peers (model 5), all 8 outcomes were significantly different (each P < .003). Interestingly, however, when we compared heterosexual-identified youths with their matched LGBTQ-identified peers (model 6), we found significant differences only on the outcomes STI testing (P = .045), perceived likelihood of STI (P = .002), and incorrect condom knowledge (P = .003).

The results of models 5 and 6 deserve further attention. Among the sample matched at the LGBTQ values (i.e., higher average victimization; model 5), LGBTQ-identified students’ odds for having anonymous sex partners, sex under the influence, and unprotected sex were 2.44, 1.85, and 2.83 times those of heterosexual-identified students, respectively (each P < .003). Among the sample matched at the heterosexual values (i.e., lower average victimization; model 6), the ORs for these outcomes were small and nonsignificant. In the model 6 sample, however, LGBTQ-identified students’ odds for thinking they would get an STI were nearly 4 times those of heterosexual-identified youths (P = .002), even though the 2 groups had statistically similar risks for many sexual risk outcomes. Also of note, LGBTQ-identified youths in both the model 5 and model 6 samples showed significantly better knowledge of condom effectiveness than did their heterosexual-identified peers (P < .003). We return to these patterns in the Discussion.

Accounting for additional forms of victimization (e.g., parental physical abuse, sexual abuse by adults, dating violence), model 7 found that an additional portion of each disparity was explained, but all risk disparities remained significant with the exception of STI testing.

To test whether the risk disparities varied between middle and high school, we reestimated models 1 through 4 and 7 with added interactions between the indicator variables for middle school and LGBTQ. In Model 1 in Table 2, all ORs were statistically significant in both middle and high school (except for unprotected sex in middle school). In addition, the ORs for number of partners and sex under the influence of drugs or alcohol were significantly higher in middle school than in high school; however, only the OR for number of sex partners remained statistically larger in middle school than in high school by model 7. Although LGBTQ-identified middle school students were much more likely than their heterosexual-identified peers to engage in sexual activities, they reported being equally likely to use protection when they engaged in sex. This is markedly different in high school, when LGBTQ-identified students were less likely to use protection than their sexually active heterosexual-identified peers. However, this difference does not appear to be attributable to lack of knowledge: both middle school and high school LGBTQ-identified students had better knowledge of condom effectiveness than did their heterosexual-identified peers. This knowledge differential was particularly pronounced in middle school.

TABLE 2—

LGBTQ–Heterosexual Odds Ratios, by Outcome, Model, and Grade Level: Wisconsin, Dane County Youth Assessment, 2008–2009

Outcome	Model 1, OR (95% CI)	Model 2, OR (95% CI)	Model 3, OR (95% CI)	Model 4, OR (95% CI)	Model 7, OR (95% CI)
Sexual activity
Grades 7–8	4.12 (2.32, 7.33)	3.49 (1.96, 6.24)	3.12 (1.66, 5.85)	2.86 (1.43, 5.72)	2.38 (1.02, 5.53)
Grades 9–12	2.42 (1.91, 3.08)	2.18 (1.68, 2.83)	2.31 (1.79, 2.97)	2.15 (1.65, 2.80)	1.65 (1.27, 2.15)
No. of partners (any)
Grades 7–8	6.06* (4.21, 8.73)	5.11* (3.19, 8.19)	4.30 (2.49, 7.43)	4.04 (2.26, 7.23)	3.88* (1.94, 7.73)
Grades 9–12	2.48 (1.83, 3.35)	2.24 (1.66, 3.02)	2.36 (1.78, 3.14)	2.18 (1.65, 2.88)	1.69 (1.29, 2.23)
No. of anonymous sex partners (any)
Grades 7–8	6.83 (3.14, 14.89)	3.74 (1.05, 13.31)	2.68 (0.78, 9.23)	2.56 (0.75, 8.71)	1.92 (0.51, 7.28)
Grades 9–12	3.39 (2.77, 4.16)	2.99 (2.32, 3.85)	2.81 (2.19, 3.59)	2.52 (1.99, 3.20)	1.80 (1.34, 2.41)
Sex under the influence of drugs or alcohol
Grades 7–8	11.13* (5.22, 23.73)	8.91* (3.45, 23.01)	6.59 (2.01, 21.65)	6.34 (1.93, 20.85)	5.40 (1.38, 21.09)
Grades 9–12	2.73 (2.00, 3.71)	2.51 (1.80, 3.50)	2.46 (1.83, 3.31)	2.39 (1.83, 3.10)	1.68 (1.20, 2.36)
Unprotected sex (among those who had sex)
Grades 7–8	1.57 (0.62, 3.99)	1.46 (0.41, 5.15)	1.32 (0.35, 5.06)	1.26 (0.34, 4.65)	1.36 (0.40, 4.66)
Grades 9–12	1.90 (1.46, 2.47)	1.91 (1.43, 2.55)	1.72 (1.24, 2.39)	1.73 (1.28, 2.34)	1.59 (1.14, 2.21)
STI testing
Grades 7–8	6.27 (2.37, 16.61)	4.58 (1.32, 15.92)	3.05 (0.73, 12.73)	2.44 (0.62, 9.56)	2.21 (0.41, 11.97)
Grades 9–12	3.32 (2.38, 4.65)	2.92 (2.06, 4.15)	2.60 (1.92, 3.53)	2.15 (1.45, 3.19)	1.55 (0.94, 2.56)
Perceived likelihood of STI
Grades 7–8	3.15 (1.14, 8.66)	2.42 (0.75, 7.80)	2.08 (0.57, 7.61)	2.26 (0.68, 7.52)	2.00 (0.45, 8.97)
Grades 9–12	2.74 (2.16, 3.49)	2.45 (1.88, 3.21)	2.12 (1.57, 2.86)	2.02 (1.55, 2.63)	1.60 (1.13, 2.27)
Incorrect condom knowledge
Grades 7–8	0.59 (0.39, 0.87)	0.66 (0.44, 1.00)	0.61 (0.42, 0.89)	0.58 (0.40, 0.85)	0.58 (0.38, 0.88)
Grades 9–12	0.80 (0.66, 0.96)	0.76 (0.61, 0.94)	0.69 (0.55, 0.86)	0.69 (0.55, 0.88)	0.70 (0.54, 0.89)

Note. CI = confidence interval; LGBTQ = lesbian, gay, bisexual, transgender, or questioning; OR = odds ratio; STI = sexually transmitted infection. Models 1 through 4 and 7 were estimated via the XTLOGIT command in Stata (StataCorp LP, College Station, TX) with school random effects and use of school-cluster jackknifed standard errors. Model 1 presents unadjusted odds ratios, including students who provided multiple-low-frequency responses on screener items. Model 2 is similar to model 1, except the multiple-low-frequency responders are removed. Model 3 conditions on race, gender, grade, age, height, weight, 2 proxies of socioeconomic status, disability status, and whether students provided 1 vs 0 low-frequency responses. Model 7 is a covariate-adjusted multilevel model, building off of model 4 and adding covariates for perceptions of family support, being kicked out of home, physical abuse by parents, sexual abuse by adults, and dating violence.

*P < .05 for LGBTQ × middle school interaction.

In our final set of analyses, we reestimated the covariate-adjusted models (models 1–4 and 7) by specific LGBTQ subgroups. Tables 1 and 3 show that there was much heterogeneity within the broader LGBTQ classification. Lesbian/gay–heterosexual and bisexual–heterosexual risk disparities tended to be largest, remaining so across nearly all models (note, though, that 4 of the 8 lesbian/gay–heterosexual disparities were statistically nonsignificant by model 7). Although more lesbian/gay–heterosexual disparities than bisexual–heterosexual disparities were nonsignificant by model 7, the ORs (Table 3) and AMEs (Table 1) were similar for most outcomes in model 7. Thus, differences between lesbian/gay–heterosexual and bisexual–heterosexual disparities in terms of statistical significance likely reflect a lower power to detect disparities between lesbian/gay and heterosexual youths than between bisexual and heterosexual youths, rather than necessarily reflecting true substantive differences between lesbian/gay- and bisexual-identified youths in terms of risk levels for these outcomes.

TABLE 3—

Sexual Risk Differentials by Outcome, LGBTQ Subgroup, and Model: Wisconsin, Dane County Youth Assessment, 2008–2009

Outcome	Model 1, OR (95% CI)	Model 2, OR (95% CI)	Model 3, OR (95% CI)	Model 4, OR (95% CI)	Model 7, OR (95% CI)
Sexual activity
Lesbian/gay	3.78* (2.40, 5.97)	3.17* (1.98, 5.09)	2.57* (1.45, 4.56)	2.73* (1.48, 5.03)	2.48* (1.38, 4.45)
Bisexual	3.92* (2.77, 5.55)	3.50* (2.44, 5.02)	3.27* (2.30, 4.64)	2.93* (1.99, 4.30)	1.89* (1.31, 2.74)
Transgender	2.28* (1.42, 3.67)	1.31 (0.67, 2.57)	1.35 (0.69, 2.64)	1.40 (0.77, 2.56)	1.20 (0.55, 2.64)
Questioning	1.41 (0.95, 2.09)	1.22 (0.80, 1.86)	1.51* (1.04, 2.18)	1.46 (0.91, 2.34)	1.32 (0.87, 2.00)
No. of partners (any)
Lesbian/gay	4.67* (2.88, 7.58)	3.85* (2.38, 6.22)	3.18* (1.60, 6.33)	3.45* (1.70, 6.99)	3.16* (1.53, 6.53)
Bisexual	3.86* (2.43, 6.13)	3.52* (2.18, 5.70)	3.33* (2.05, 5.40)	2.98* (1.77, 5.01)	1.96* (1.16, 3.31)
Transgender	3.24* (1.83, 5.73)	1.82 (0.92, 3.62)	1.97 (0.83, 4.66)	2.06 (0.89, 4.77)	1.82 (0.60, 5.46)
Questioning	1.37 (0.83, 2.26)	1.11 (0.68, 1.81)	1.35 (0.89, 2.03)	1.25 (0.79, 2.00)	1.20 (0.71, 2.02)
No. of anonymous-sex partners (any)
Lesbian/gay	6.42* (4.16, 9.89)	4.57* (2.79, 7.50)	3.40* (1.90, 6.08)	3.76* (2.06, 6.87)	2.89* (1.50, 5.56)
Bisexual	4.61* (3.43, 6.19)	4.16* (2.94, 5.90)	3.64* (2.54, 5.21)	3.23* (2.23, 4.68)	1.97* (1.26, 3.07)
Transgender	5.77* (3.40, 9.80)	2.41* (1.01, 5.73)	1.93 (0.77, 4.81)	1.98 (0.82, 4.82)	1.57 (0.47, 5.27)
Questioning	2.75* (1.53, 4.94)	1.99* (1.16, 3.43)	2.30* (1.49, 3.54)	2.01* (1.22, 3.30)	1.67 (0.96, 2.90)
Sex under the influence of drugs or alcohol
Lesbian/gay	4.28* (2.18, 8.42)	3.16* (1.52, 6.54)	2.22 (0.78, 6.32)	2.71 (0.89, 8.25)	2.04 (0.60, 6.91)
Bisexual	4.32* (2.93, 6.37)	4.00* (2.56, 6.24)	3.54* (2.37, 5.27)	3.52*(2.32, 5.33)	2.15* (1.35, 3.43)
Transgender	4.53* (2.80, 7.35)	2.24 (0.98, 5.13)	2.17 (0.81, 5.84)	2.20 (0.87, 5.56)	1.68 (0.45, 6.24)
Questioning	1.76 (0.98, 3.15)	1.29 (0.70, 2.38)	1.59* (1.04, 2.42)	1.51* (1.01, 2.26)	1.38 (0.82, 2.33)
Unprotected sex (among those who had sex)
Lesbian/gay	3.87* (1.72, 8.72)	4.31* (1.52, 12.26)	3.92* (1.29, 11.88)	4.06* (1.33, 12.35)	3.44* (1.12, 10.64)
Bisexual	1.86* (1.24, 2.79)	1.91* (1.25, 2.90)	1.61* (1.00, 2.60)	1.70* (1.05, 2.74)	1.52 (0.90, 2.56)
Transgender	1.41 (0.66, 3.04)	0.71 (0.19, 2.64)	0.80 (0.22, 2.95)	0.77 (0.23, 2.61)	0.71 (0.20, 2.50)
Questioning	2.15* (1.53, 3.03)	1.74* (1.20, 2.54)	1.56 (0.96, 2.55)	1.66* (1.00, 2.75)	1.60 (0.99, 2.60)
STI testing
Lesbian/gay	5.53* (3.39, 9.03)	3.84* (1.99, 7.43)	2.72* (1.24, 5.93)	2.38 (0.97, 5.83)	2.15 (0.92, 5.02)
Bisexual	5.30* (3.33, 8.45)	4.72* (2.96, 7.55)	3.57* (2.60, 4.91)	2.84* (1.89, 4.26)	1.71* (1.01, 2.90)
Transgender	3.38* (1.79, 6.39)	1.82 (0.61, 5.45)	2.06 (0.59, 7.12)	2.28 (0.71, 7.32)	1.75 (0.40, 7.68)
Questioning	1.15 (0.45, 2.91)	0.86 (0.35, 2.11)	0.89 (0.42, 1.88)	0.75 (0.33, 1.73)	0.65 (0.22, 1.90)
Perceived likelihood of STI
Lesbian/gay	3.65* (1.77, 7.52)	2.63* (1.20, 5.76)	2.34* (1.09, 5.01)	2.31 (1.00, 5.36)	1.85 (0.76, 4.51)
Bisexual	2.76* (2.22, 3.44)	2.46* (1.83, 3.31)	2.04* (1.46, 2.84)	2.11* (1.48, 3.00)	1.69* (1.11, 2.56)
Transgender	3.74* (1.69, 8.28)	1.75 (0.50, 6.15)	1.67 (0.36, 7.87)	1.55 (0.34, 6.99)	1.34 (0.25, 7.20)
Questioning	3.09* (2.06, 4.63)	2.43* (1.48, 3.97)	2.03* (1.19, 3.48)	1.80* (1.03, 3.15)	1.31 (0.58, 2.96)
Incorrect condom knowledge
Lesbian/gay	0.75 (0.42, 1.34)	0.64 (0.36, 1.13)	0.73 (0.37, 1.45)	0.71 (0.34, 1.46)	0.72 (0.36, 1.44)
Bisexual	0.61* (0.47, 0.79)	0.60* (0.45, 0.79)	0.61* (0.47, 0.79)	0.60* (0.45, 0.80)	0.63* (0.46, 0.88)
Transgender	0.80 (0.50, 1.29)	0.94 (0.56, 1.57)	0.84 (0.47, 1.51)	0.83 (0.46, 1.51)	0.79 (0.44, 1.43)
Questioning	0.89 (0.59, 1.35)	0.85 (0.57, 1.25)	0.71 (0.49, 1.02)	0.70 (0.47, 1.04)	0.66* (0.48, 0.91)

Note. CI = confidence interval; LGBTQ = lesbian, gay, bisexual, transgender, or questioning; OR = odds ratio; STI = sexually transmitted infection. Models 1 through 4 and 7 were estimated via the XTLOGIT command in Stata (StataCorp LP, College Station, TX) with school random effects and use of school-cluster jackknifed standard errors. Model 1 presents unadjusted odds ratios, including students who provided multiple-low-frequency responses on screener items. Model 2 is similar to model 1, except the multiple-low-frequency responders are removed. Model 3 conditions on race, gender, grade, age, height, weight, 2 proxies of socioeconomic status, disability status, and whether students provided 1 vs 0 low-frequency responses. Model 7 is a covariate-adjusted multilevel model, building off of model 4 and adding covariates for perceptions of family support, being kicked out of home, physical abuse by parents, sexual abuse by adults, and dating violence.

*P < .05.

By comparison, questioning–heterosexual risk disparities were much smaller. Transgender–heterosexual risk disparities were large and significant in model 1 (which included multiple-low-frequency responders), but decreased to nonsignificant levels across all models and outcomes once we excluded multiple-low-frequency responders. The conspicuous change in the transgender–heterosexual risk disparities from the model that contained the multiple-low-frequency responders (i.e., model 1) to the models that excluded the multiple-low-frequency responders raised concerns about the validity of the transgender-identified sample. As a result, we reran models 1 through 4 and 7, excluding students who identified as transgender only, to verify that the patterns of LGBTQ–heterosexual risk disparities were not driven by this potentially invalid set of responders. These supplemental analyses verified that LGBQ–heterosexual risks were indeed significant for all outcomes and of generally the same magnitude as the LGBTQ–heterosexual disparities. Thus, the transgender-only-identified group was not driving the results.

DISCUSSION

This population-based study of middle and high school students is the first such study to estimate LGBTQ–heterosexual sexual risk disparities through a process of screening out potentially mischievous responders, developing propensity-score-matched estimates, and examining a variety of victimization factors in a single sample. Our findings provide evidence that LGBTQ–heterosexual sexual risk disparities tend to begin early (by middle school) and that they vary across the LGBTQ subgroups. In addition, peer victimization does indeed account for a portion of the disparities, consistent with previous studies that employed different methodologies.^2,5 Our results also suggest that accounting for childhood physical and sexual abuse,^2,18,29 dating violence,¹⁹ and family rejection¹⁷ explains an additional portion of the disparities. Thus, our findings are consistent with other studies suggesting that victimization explains part of LGBTQ–heterosexual risk disparities. Importantly, however, substantial risk disparities remained even after we accounted for various forms of victimization, which suggests that models exploring the sources of sexual risk disparities between LGBTQ and heterosexual youths should consider a broader range of factors than the forms of victimization included in the present study.³⁰

Perhaps most intriguing, the estimates based on propensity score matching suggest that LGBTQ–heterosexual risk disparities are more often significant among the matched sample of youths who are victimized more often (model 5); however, differences in self-perceptions of STI risk were substantial and significant among the sample with lower victimization despite the comparable sexual risk-taking patterns of LGBTQ- and heterosexual-identified youths in that sample (model 6). What is the source of this discrepancy? A hint as to an explanation may be found in our recent study using the same sample and analytic design,³⁰ where LGBTQ–heterosexual disparities for suicide ideation and suicide attempts were significant in both matched samples (i.e., models 5 and 6). The combined findings of these studies suggest that, among the higher-victimization sample, LGBTQ-identified youths are at greater risk for both sexual risk-taking and suicide than their matched heterosexual-identified peers; however, among the lower-victimization sample, LGBTQ-identified youths are at greater risk for suicide and for perceptions of getting an STI, but are generally not at greater risk for engaging in sexually risky activities (e.g., anonymous sex partners, sex under the influence of drugs or alcohol). This finding raises important questions about the messages LGBTQ-identified youths receive that may contribute to heightened thoughts of suicide and a perception that STIs are in their future, even when their rates of engaging in sexually risky behavior are similar to those of their heterosexual-identified peers and they experience relatively low levels of victimization.

Although precautions were taken to ensure the reliability of the data and assess the stability of the estimates, several limitations should be noted. First, the data were self-reported, and we have noted that some students may have willfully provided mischievous responses that can skew the results.³¹ To mitigate this, we employed a screening technique to identify and exclude potentially mischievous responders.⁴ Second, the sexual-orientation question on the DCYA does not have “heterosexual” as an option; thus, some students who chose “none of the above” may not have been heterosexual-identified (e.g., they may have been asexual-identified or queer-identified). In addition, the question combines “transgender” (a gender identity) with sexual-orientation identities; thus, a youth who was transgender-identified and heterosexual-identified could not be identified in our sample and would only be considered as transgender-identified. To remedy these concerns in the future, we have recommended to the DCYA that on future iterations they (1) include heterosexual as an explicit option and (2) separate out questions about sexual orientation and gender identity.

Third, our method for identifying LGBTQ youths is based on a single question related to identification, rather than on a more comprehensive measure incorporating same-sex attraction and same-sex sexual behavior.^32,33 Upon our recommendation, the DCYA now assesses same-sex attraction. Fourth, the data were collected in a relatively progressive county (one that, at the time of this survey, was represented by the only openly lesbian US congresswoman), and thus the risk disparities may be smaller in our sample than in less progressive areas. Although this may restrict generalizations, it should also be noted that this implies that the sample composition may be biased against finding differences. Fifth, the data are cross-sectional and do not facilitate causal inferences with regard to the effects of victimization on sexual risk outcomes. Finally, although we could study heterogeneity within the LGBTQ group, there were other dimensions of heterogeneity that we could not study because of small cell sizes. For example, sexual-minority youths who are also members of a racial/ethnic minority may be particularly at risk,³⁴ but our sample size was too small for such comparisons.

Despite these limitations, our results add new insights to the role of victimization in explaining LGBTQ–heterosexual risk disparities. As previous accounts have suggested, victimization does explain part of the disparities; however, it does not explain their full extent. In addition, peer victimization may play a smaller role in explaining LGBTQ–heterosexual disparities in outcomes related to suicide³⁰ and STIs, as suggested by the fact that the ORs in these outcomes were similar at higher (model 5) and lower (model 6) levels of peer victimization. Speculatively, these disparities may result in part from stigmatizing messages regarding sexual minorities that youths receive, which may create psychological distress in any LGBTQ-identified child, regardless of how much bullying they are exposed to.^35–37 By contrast, peer victimization may play a larger role in explaining disparities in sexual behavior (e.g., sex under the influence of drugs or alcohol), as suggested by the fact that disparities were found at higher, but not lower, levels of peer victimization. That is, only when LGBTQ-identified youths are victimized often do we see differences in sexually risky behavior between them and heterosexual-identified youths.

Our findings suggest that addressing peer victimization may help reduce some disparities, but other interventions may be necessary, such as those that alter the school climate in regard to LGBTQ issues.^{6,13,24,35–37} Previous research has suggested that staff training on the diversity of family types (e.g., nonheterosexual parents)³⁸ and on addressing LGBT harassment^39,40 might help alter the school climate, as might gay–straight alliances and the incorporation of LGBTQ issues into curricula.^41,42 Using rigorous methods and multiple modeling strategies, our research provides new evidence that victimization (and broader forms of abuse) explain part of LGBTQ–heterosexual disparities in sexual risk in middle and high school; however, disparities remain—across all modeling strategies—suggesting that at least part of the explanation for these disparities lies beyond the traditional set of individual-level victimization and abuse factors.