Introduction

Most population-level studies of the relationship between gender and health evaluate differences between two groups: men and women, typically differentiated through biological, social, and behavioral characteristics (Case and Paxson 2005; Courtenay 2000; Gorman and Read 2006; Rogers et al. 2010; Udry 1994). However, scholars often conceptualize characteristics of men and women using distinct theoretical categorizations. Sex is typically assigned at birth as male or female based on physical characteristics, whereas gender consists of social identities and lived experiences (Butler 1986; West and Zimmerman 1987), including the traditional categories of men and women, and boys and girls. Individuals’ gender identities may or may not correspond to the gender others attribute to their birth-assigned sex or physical characteristics (Krieger 2003). Moreover, sex and physical characteristics at birth do not always reflect binary classifications (Fausto-Sterling 2000), and aspects related to them may change during the life course. If this theoretical distinction is truly meaningful, research that compares only men versus women may fail to account for population-level health patterns that arise from social expectations or experiences linked to gender identity.

Most large-scale surveys do not measure sex and gender distinctly from each other. These surveys usually code respondents as simply male or female, conflating sex and gender as one and the same (Reisner et al. 2015a, 2015b). This practice reflects common shortcomings in large-scale survey data, which rarely capture gender identities with the same level of precision emphasized in theoretical work (Compton 2015; Dinno et al. 2014; Schilt and Lagos 2017; Westbrook and Saperstein 2015). However, it is possible to identify and use certain existing measurements as proxies for biological sex and lived gender identities using population-level data, including by identifying transgender and gender-nonconforming respondents.

Since 2014, the Centers for Disease Control (CDC) Behavioral Risk Factor Surveillance System (BRFSS) has included a series of questions that ask whether a respondent is transgender, gender-nonconforming, or cisgender (nontransgender) (Baker and Hughes 2016; CDC 2014–2016a; Herman 2014). The term transgender refers to individuals who were assigned a sex at birth that does not correspond to their lived gender identity. A transgender woman, for example, was assigned male at birth but identifies as a woman; a transgender man was assigned female at birth but identifies as a man. Gender-nonconforming refers to individuals who do not exclusively identify as either men or women, regardless of their sex assigned at birth (male or female). Gender-nonconforming identities can intersect with various expressions of cisgender or transgender identity (Miller and Grollman 2015), but the BRFSS does not ask respondents whether they identify as gender-nonconforming unless they identified as transgender during the interview, and the survey does not allow gender-nonconforming individuals to select multiple gender identities that may reflect their experience (CDC 2014-2016b; also see the online appendix, section A). Therefore, this study makes claims only about those gender-nonconforming respondents who explicitly identify as transgender and who identify more closely with a gender-nonconforming identity than with an identity as a transgender man or woman. Cisgender refers to individuals who identify as men or women and were assigned a sex at birth that corresponds to their lived identity. A cisgender man was assigned male at birth and identifies as a man; a cisgender woman was assigned female at birth and identifies as a woman.

A person’s relationship to prevailing gender roles, and the social expectations that come with them, may shape social opportunities and challenges that influence health in ways that are not primarily associated with one’s sex assignment at birth. This study examines whether overall self-rated health differs significantly among five distinct gender identity groups: cisgender men, cisgender women, transgender women, transgender men, and gender-nonconforming individuals. These groups represent intersecting axes of sex (male, female, and so on), gender identity (man, woman, gender-nonconforming), and whether one is transgender or cisgender (see Fig. 1).

Fig. 1
figure 1

Conceptual diagram, sex and gender identity categories captured or inferable from Behavioral Risk Factor Surveillance System surveys from 2014–2016. Sex categories are outlined with solid lines; gender identity categories are outlined with dashes

Full size image

Transgender and gender-nonconforming individuals face pervasive social and health-related disadvantages (Bockting et al. 2013; Bradford et al. 2013; Brennan et al. 2012; Connell 2010; Hughto et al. 2015; James et al. 2016; Lombardi 2009; Schilt 2010), but few studies have corroborated these findings with samples that are not based on convenience or community-based recruitment methods until recently. Research based on a population sample from Massachusetts suggests that the patterns identified in convenience samples do not reflect a transgender health disadvantage at the population level (Conron et al. 2012), but these findings are limited in generalizability because they rely on data collected from a single U.S. state. Research based on the first multistate sample of the BRFSS in 2014 has found that transgender and gender-nonconforming individuals, evaluated together as a broad group, have higher odds of reporting poor physical and mental health when compared with cisgender individuals (Meyer et al. 2017) but no significant difference in alcohol consumption or rates of breast cancer screening (Blosnich et al. 2017; Narayan et al. 2017). Although the 2014 BRFSS’ larger sample and more diverse sampling frame of 19 states inspires more confidence, these studies did not include enough observations of transgender and gender-nonconforming respondents (n = 691) to compare subpopulation health patterns among transgender women (n = 363), transgender men (n = 212), and gender-nonconforming respondents (n = 116). Newer studies have used a slightly larger 27-state sample based on pooling the 2014 and 2015 BRFSS, which included 724 transgender women, 449 transgender men, and 270 gender-nonconforming adults. These studies have been able to establish some insights into differences in barriers to health care and HIV testing among transgender men, transgender women, and gender-nonconforming adults, but they did not use these categories to compare differences in overall health among these groups (Gonzalez and Henning-Smith 2017; Pitasi et al. 2017). In this study, I leverage a much larger sample pooled from 31 U.S. states and one U.S. territory in 2014, 2015, and 2016 that includes 1,075 transgender women, 699 transgender men, and 450 gender-nonconforming respondents.

Do Gender Identity and Sex Contribute Differently to Health?

Gender role expectations for men and women vary across different cultures and socioeconomic strata, and they change over time (Furtado et al. 2013; Kalmijn 2013). This suggests that gender roles are not uniformly predetermined by biology even among the cisgender majority, despite strong patterns of similarity across contexts. Nevertheless, population scholars have tended to assume that deviation from gender norms stems from concrete conflicts within a delimited set of external goals, such as career progress (Park et al. 2015; Watkins 1993). Few of these studies have considered that deviation from gender norms may stem from an individual’s wholescale conflict with the gendered identity that they are expected to adopt based on their sex rather than only a few points of tension. This results in work that overemphasizes slight variations in gender role patterns that nevertheless remain firmly rooted in an overarching preference for gender conformity, largely ignoring transgender and gender-nonconforming identities (Deutsch 2007; Miller and Grollman 2015; Risman 2009).

Indeed, wholescale tensions between individuals’ gender identities and social expectations based on sex may shape health outcomes through minority stress processes (Meyer 1995), in which social marginalization can produce elevated levels of psychosomatic stress. In a large number of social settings—such as public restrooms, medical appointments, and educational institutions—transgender and gender-nonconforming individuals often experience forms of stigma, discrimination, and alienation that can lead to elevated levels of psychological distress and poor health (Herman 2013; Miller and Grollman 2015; Reisner et al. 2015a, 2015b; shuster 2016). These sources of stress can manifest in a variety of ways among transgender and gender-nonconforming individuals, including through experiences of gender dysphoria, anti-transgender violence, and complications related to seeking medical treatment (Lombardi 2009). With these patterns of inequality in mind, I formulate five hypotheses (online appendix, Table S1). I begin by hypothesizing an overall health disadvantage for transgender and gender-nonconforming respondents in contrast to cisgender respondents:

  • Hypothesis 1: Transgender and gender-nonconforming respondents will have higher odds of reporting poor health than will cisgender respondents.

Some scholars have contended that differences between men and women are largely explained by biological and genetic factors unique to each population rather than by socially driven factors related to gender roles and identity (Udry 1994). Studies focusing on sex-based health differences have found that cisgender women generally live longer than cisgender men but that they experience worse self-rated health and higher incidence of hospitalization later in life (Case and Paxson 2005; Gorman and Read 2006). The picture of how sex influences health is less clear when comparing transgender women, transgender men, and gender-nonconforming individuals. To evaluate biological and genetic factors versus social factors, I generate a second hypothesis that examines differences in the odds of reporting poor self-rated health between individuals who were presumably assigned female or male at birth. This hypothesis evaluates whether there are detectable disparities in self-reported health that correspond to sex assigned at birth rather than a common identity as men or women:

  • Hypothesis 2: Respondents who were assigned female at birth (cisgender women and transgender men) will have higher odds of reporting poor health than will respondents who were assigned male at birth (cisgender men and transgender women).

Differences in behavioral expectations for men and women regarding household roles, education, seeking medical treatment, and smoking shape health disparities between cisgender men and women (Courtenay 2000; Rogers et al. 2010; Ross et al. 2012; Saltonstall 1993; Springer and Mouzon 2011; Stroope 2015). It is difficult to predict how closely patterns of behavioral and social difference, such as differences in smoking rates and marital status between transgender men and women, will resemble established differences between cisgender men and women. Many health stressors may have different degrees of salience for transgender men versus transgender women, including risk patterns for anti-transgender harassment and violence, complications due to different hormonal and surgical processes related to transitions (Bockting et al. 2013; Lombardi 2009), and different forms of discrimination in daily life (Schilt 2010). Transgender women have higher risk of experiencing psychosocial distress and HIV-positive status than transgender men (Brennan et al. 2012; Clements-Nolle et al. 2001). Transgender men may experience more discrimination in health settings and delayed access to medical procedures compared with transgender women (Bradford et al. 2013; Grant et al. 2011; James et al. 2016). Other studies have found no statistically significant difference in instances of transgender-related hostility between transgender men and women (Bockting et al. 2013; Lombardi 2009). Hypothesis 3 thus evaluates gender identity–based differences between men and women, both cisgender and transgender:

  • Hypothesis 3: Cisgender women will have higher odds of reporting poor health than cisgender men, and transgender women will have higher odds of reporting poor health than transgender men.

Research on gender-nonconforming individuals suggests that they are more likely to be disadvantaged than transgender men and women because of the social costs of not fitting into a commonly recognized gender category (Connell 2009; Harrison et al. 2012; James et al. 2016; Lombardi 2009; Schilt and Westbrook 2009). Many social settings and structures assume distinct and rigid roles for men and women, including transgender men and women, such as the workplace (Connell 2010; Schilt 2010), the doctor’s office (Bradford et al. 2013; shuster 2016), and the law (Meadow 2010; Westbrook and Schilt 2014). Even if some gender-nonconforming individuals may have more success than transgender men or women in concealing their identities from stigma, gender nonconformity can lead to diminished social status, higher rates of violence (Schilt and Westbrook 2009), and social isolation (Westbrook and Saperstein 2015). Although transgender men and women are more likely to not have health insurance, gender-nonconforming adults are more likely to have skipped medical treatment because of cost (Gonzalez and Henning-Smith 2017). Some evidence suggests that gender-nonconforming individuals experience poorer overall health patterns than transgender men and women (Miller and Grollman 2015), but this relationship remains understudied at the population level using nonconvenience sampling methods (Institute of Medicine 2011; Schilt and Lagos 2017). Hypothesis 4 thus evaluates the relevance of identifying as gender-nonconforming to overall health:

  • Hypothesis 4: Gender-nonconforming respondents will have the highest odds of reporting poor health of any group.

Gender identities, and their influence on health, do not occur in isolation from other important social factors, such as age, race/ethnicity, socioeconomic status (SES), social support, and behavioral factors. Gender differences in mortality often differ widely by age cohort—for example, because of differences in behavioral risk factors among men versus women in a given cohort (Preston and Wang 2006). Racial disparities in overall health among black, Hispanic, and white men are more pronounced than those among black, Hispanic, and white women (Gorman and Read 2006; Umberson et al. 2014). Education, which often corresponds to broader socioeconomic patterns, has a larger association with self-rated health for women than for men (Ross et al. 2012), and lower levels of poverty among women play a significant role in the gender health gap (Rogers et al. 2010). Marriage, as a form of social support, is associated with a health advantage (Rendall et al. 2011; Waite 1995), whereas divorce and separation often lead to loss of health insurance coverage for lower-SES women (Peters et al. 2014). Smoking rates are typically higher among men than among women, but they may be narrowing for more recent cohorts as gendered behavioral trends become more egalitarian (Case and Paxson 2005; Preston and Wang 2006; Rogers et al. 2010).

Evidence for whether these gender differences are reflected among transgender and gender-nonconforming populations is sparse, but different patterns between these groups can be inferred from existing research. Elderly transgender adults report poorer overall health than elderly cisgender adults, perhaps because of the greater stigma and lack of social support faced by elderly transgender individuals (Frederiksen-Goldsen et al. 2014). Compared with white and wealthier transgender individuals, black, Hispanic, and lower-SES transgender individuals face more transgender-related stigma and discrimination (Lombardi 2009) as well as higher risk of HIV and sexually transmitted infections (Nuttbrock et al. 2009; Sevelius 2013). Transgender and gender-nonconforming children, adolescents, and young adults may face heavy discrimination and hostility in educational settings, from primary school through college (Bradford et al. 2013; Reisner et al. 2015a, 2015b), which could lead to differences in educational attainment and socioeconomic positions later in life. Transgender men report higher smoking rates than transgender women, and gender-nonconforming individuals have a higher smoking rate than any other group (Miller and Grollman 2015). Very little is known regarding marriage patterns, particularly their implications for health patterns, among transgender and gender-nonconforming populations (Biblarz and Savci 2010). However, convenience-based studies have suggested that transgender individuals have lower rates of marriage than the general U.S. population (James et al. 2016). Adjusting for socioeconomic factors situates comparisons between gender identity groups in a broader social context:

  • Hypothesis 5: Adjustments for SES, marital status, and smoking rates will reduce the overall risk of reporting poor health for all groups, but transgender and gender-nonconforming individuals will still face significantly higher odds of reporting poor health than cisgender men.

Data and Methods

Data used in this study come from the CDC Behavioral Risk Factor Surveillance System (BRFSS) between 2014 and 2016, gathered from 31 states and one U.S. territory (see online appendix, Table S2). The BRFSS is a nationwide health survey of noninstitutionalized U.S. adults conducted by each U.S. state and territory’s public health department. The survey uses household-based probability sampling and random digit dialing of landlines and cellular phones to recruit respondents. Since 2014, the BRFSS has adopted a standardized sexual orientation and gender identity module that asks respondents whether they identify as transgender and, if so, whether they identify as male-to-female transgender (transgender women), female-to-male transgender (transgender men), or gender-nonconforming (see online appendix, section A). This module is optional, and 32 U.S. state and territory health departments have implemented it in all or some administrations of the BRFSS since 2014, producing a final analytic sample that includes 598,286 respondents (see online appendix Table S2). Pooling the data in this manner greatly expands the breadth and representability of previous samples on transgender health, and it provides more information about health-related trends in the transgender population than what is available through convenience samples. Survey weighting ensures that analyses of the pooled sample are adjusted to reflect the state’s actual population and the number of years of data available from each state.

Measures

Following prior demographic work on intercategorical gender differences in self-rated health (Gorman et al. 2015), I use a dichotomous outcome classification of self-rated health as the dependent variable. “Poor” and “fair” self-rated health are combined as one outcome, categorized as “poor health.” “Good,” “very good,” and “excellent” self-rated health are combined as another outcome, categorized as “good health.” Self-rated health is a robust predictor of mortality in populations (DeSalvo et al. 2006; Idler and Benyamini 1997; Jylhä 2009), irrespective of SES (Frankenberg and Jones 2004; Gorman and Sivanganesan 2007; Quesnel-Valée 2007). I also evaluate potential factors that may indicate stressors and may intersect with gender identity, such as SES, race/ethnicity, marital status, and smoking (Denney et al. 2013; Rendall et al. 2011; Waite 1995). Because self-rated health is in the core module of the BRFSS asked by every state public health agency, very few cases are missing answers to this question (353, or <0.01 %); the few cases missing answers are excluded from regression models and predicted probabilities.

In the BRFSS, the following questions are used to collect information on gender identity. First, the survey interviewer asks, “Do you consider yourself to be transgender?” If the respondent answers “No,” the survey administrator moves on to other questions. If the respondent answers “Yes,” the administrator asks, “Do you consider yourself to be male-to-female, female-to-male, or gender-nonconforming?” If a respondent needs help defining “transgender” or “gender-nonconforming” to understand the questions, interviewers provide a definition according to a uniform script (see online appendix, section A). Answers to this question are combined to produce unweighted samples of cisgender men (n = 218,362), cisgender women (n = 298,391), transgender women (n = 1,078), transgender men (n = 701), and gender-nonconforming respondents (n = 450).

The ability of the BRFSS’ gender identity questions to provide information about respondents’ sex has recently come under scrutiny when identifying transgender respondents because survey interviewers initially assess respondents’ sex based on their interpretation of the timbre of a respondent’s voice (Riley et al. 2017). However, the main concerns that arise from this method correspond to the accuracy of findings regarding sex-specific medical tests, such as exams for prostate cancer. The majority of research based on the BRFSS considers this measure of sex to be sufficient for studying sex-based differences among cisgender populations. By definition, cisgender individuals identify with a gender that corresponds to the sex they were assigned at birth. Although sex is indeed a multidimensional trait combining multiple physical characteristics, as well as various medical and legal classifications (Fausto-Sterling 2000), it is possible to assume that in the vast majority of cases, cisgender women were assigned female at birth, and cisgender men were assigned male at birth. Along the same lines, it is possible to infer that people who identify as transgender men were assigned female at birth and that people who identify as transgender women were assigned male at birth, based on the definitions of these terms provided to respondents by the interviewers (see online appendix, section A).

These limitations also have implication for findings based on gender-nonconforming respondents. Data on gender-nonconforming respondents can be used to analyze the overall influence of gender-nonconforming identity, but they do not offer insight into the relative role of specifically male or female sex assignment at birth. Furthermore, as mentioned earlier, the BRFSS does not ask gender-nonconforming respondents whether they also identify as a transgender man or transgender woman even though these identities often overlap, and the survey does not allow respondents to identify as gender-nonconforming if they do not first tell the interviewer that they identify as transgender. Among the sample, 79,657 respondents (13.3 %) refused to answer, did not know, or were not asked the question regarding their gender identity. For a relatively recently added question, this degree of missingness is to be expected (Little and Rubin 2014), particularly given the level of stigma associated with transgender and gender-nonconforming identities. These missing cases are imputed in all estimations, but results do not vary significantly from those based on models that exclude missing observations.

To account for the influence of aging on self-rated health, I include being 65 years old or older at the time of interview as a dichotomous measure. I also account for racial and ethnic categories, divided into the following groups: non-Hispanic whites, non-Hispanic blacks, Hispanics of any race, and non-Hispanic respondents who are multiracial or identify with a race or ethnicity not encompassed by the preceding categories. SES is evaluated through level of education attained (no high school diploma, having a high school diploma or GED, having attended college). These distinctions correspond to distinct categorical differences in lifetime earnings (Kane and Rouse 1995; Tyler and Lofstrom 2009). In addition to education, I use respondents’ income to account for the broader relationship between general access to material resources and health (Benzeval and Judge 2001; Deaton and Paxson 1998); this is a dichotomous measure of whether the respondent’s household income is greater than $50,000. To account for the number of potential earners and dependents living on this level of income, I also include two measures for any children or any other adults living in the respondent’s household in all analyses that include income.

I use respondents’ marital status and history as a measure of social support because marriage has well-established links to overall health (Rendall et al. 2011; Waite 1995). Respondents are classified as being currently married or in an unmarried couple, having ever been married (including respondents who are widowed, divorced, or separated), and having never been married. BRFSS surveys occasionally collect measures of overall emotional support, which may come from relationships outside marriage and partnership, but only a few states administer that module at this time. I thus do not use those measures here, even though they may be useful for evaluating whether transgender and gender-nonconforming individuals benefit from different forms of social support than does the general population (Pfeffer 2012). Finally, I consider smoking as a behavioral risk factor that often differs between men and women (Preston and Wang 2006) as well as between transgender and cisgender populations in both patterns of use and marketing (Amos et al. 2012; Smith et al. 2007); it is measured through groupings of respondents who are current smokers, are former smokers, or have never smoked.

Analysis

To optimize each step of the data preparation and analysis process with the best computational resources available, I use both R and Stata. Data are pooled and recoded using R 3.4.0. Then, I impute missing data using an expectation-maximization with bootstrapping (EMB) algorithm using the Amelia package, version 1.7.4 in R (Honaker et al. 2011), using a total of five imputations, as recommended by the program developers. Self-rated health is included in the imputation model, but after imputation, I delete observations that originally had missing outcome variables, following the “multiple imputation, then deletion” (MID) process (Gorman et al. 2015; Von Hippel 2007).

For the estimation of models, I combine multiple imputation estimations with survey weighting (using first-order Taylor linear approximation), which accounts for the complex design of the BRFSS, using the mi and svy functions in Stata 15.0 (StataCorp 2017). I evaluate odds ratios and confidence intervals obtained from logistic regression models to examine whether particular factors are associated with disparities in self-rated health. Logistic regression is suitable for analyzing binary outcomes, but comparing odds ratios across models can be problematic (Mood 2010). To compare gender identity groups with each other individually, I use logistic regression parameters to calculate predicted probabilities; this produces mean estimates for the probability of reporting poor health, comparing the baseline model with the fully adjusted model (Bartus 2005; Gorman et al. 2015). The mimargins function in Stata generates predicted probabilities, which can be used to make pairwise comparisons based on the average marginal effect (AME) of belonging to each gender identity group. All significance tests are evaluated by applying the Benjamini-Hochberg control for the false discovery rate (0.05) for multiple comparisons (Benjamini and Hochberg 2000).

Results

Sample Characteristics

Table 1 presents characteristics of the entire analytic sample. Table 2 presents bivariate relationships between gender identity and all sample characteristics, showing that gender-nonconforming respondents have the highest prevalence of reporting poor health (30.28 %), followed by transgender men (23.32 %), transgender women (18.42 %), cisgender women (18.16 %), and finally cisgender men (16.64 %). As a group, cisgender respondents have lower overall rates of poor self-reported health than transgender and gender-nonconforming respondents. Cisgender women have a higher prevalence of poor self-reported health than cisgender men, and transgender men have a higher prevalence than transgender women. Gender-nonconforming adults have the highest proportion of respondents reporting poor health of any group. Transgender men, transgender women, and gender-nonconforming respondents have lower proportions of white respondents than cisgender men and women, with higher proportions of black, Hispanic, and other ethnic/racial groups. Compared with transgender men and women, gender-nonconforming respondents have higher rates of college attendance and yearly household incomes more than $55,000, and far lower rates of not having completed high school. Gender-nonconforming respondents also have the lowest rate of ever having smoked compared with all other groups.

Table 1 Descriptive overview: Survey-weighted percentages from analytic sample of Behavioral Risk Factor Surveillance System, 2014–2016
Full size table
Table 2 Sample characteristics (percentages): Pooled analytic sample from Behavioral Risk Factor Surveillance System surveys, 2014–2016
Full size table

Logistic Regression Models for Poor Self-rated Health

Table 3 presents the odds ratios from a series of logistic regression models predicting self-reported poor health. All tests that are rejected in the original models are also rejected after applying the Benjamini-Hochberg adjustment (Benjamini and Hochberg 2000) and are indicated in the shaded cells. Model 1 compares the cisgender and transgender samples. As a group, the transgender sample has a significant overall health disadvantage compared with the cisgender group (odds ratio (OR) = 1.37, 95 % confidence interval (CI) = 1.14,1.63), replicating the findings of past research (Meyer et al. 2017) and confirming Hypothesis 1. Model 2 compares each gender identity group separately in reference to cisgender men. The odds of reporting poor health among transgender women do not differ significantly from cisgender men, but gender-nonconforming respondents (OR = 2.18, CI = 1.51,3.15), transgender men (OR = 1.53, CI = 1.11,2.11), and cisgender women (OR = 1.11, CI = 1.08,1.14) report significantly higher odds of reporting poor health than cisgender men.

Table 3 Estimated odds ratios from logistic regression predicting poor self-reported health: Pooled analytic sample from Behavioral Risk Factor Surveillance System surveys, 2014–2016
Full size table

When controls for age and race/ethnicity are included in Model 3, cisgender women (OR = 1.08, CI = 1.05,1.11), transgender men (OR =1.45, CI = 1.05,2.01), and gender-nonconforming respondents (OR = 2.10, CI = 1.41,3.13) retain a statistically significant overall health disadvantage in comparison with cisgender men. Ultimately, there are no major differences in direction or significance between Models 2 and 3, and the odds of reporting poor health remain insignificant among transgender women in reference to cisgender men. Model 4 further adjusts for socioeconomic and behavioral factors, including levels of education and income, household size, marital status, and lifetime smoking history. Cisgender women (OR = 1.09, CI = 1.05,1.13) and gender-nonconforming respondents (OR = 2.06, CI = 1.25,3.40) continue to have statistically significant higher odds of poor self-rated health in comparison with cisgender men. However, transgender men no longer display any statistically significant differences in comparison with cisgender men in this model, and neither do transgender women.

A central finding here is that a discernable disadvantage persists among the gender-nonconforming subpopulation: their odds of reporting poor health remain more than twice as high as those of cisgender men even after I adjust for demographic, socioeconomic, and behavioral factors. Even in the fully adjusted model, the association between gender-nonconforming identity and reporting poor health is stronger than the association between poor self-rated health and being a current smoker (OR = 1.85, CI = 1.77,1.93). Another significant finding is that there is no discernably significant disadvantage among transgender men after I adjust for key demographic, socioeconomic, and behavioral factors in Model 4. Transgender men have the lowest proportions of having attended college or having an income greater than $55,000 per year, the highest proportion of having any children in their households, as well as the highest rate of current smokers of any group, suggesting that these factors are closely related to their overall health (see Table 2). Surprisingly, in comparison with cisgender men, transgender women are not disadvantaged in terms of their self-rated health.

To make comparisons between the gender identity groups, Fig. 2 presents the average predicted probabilities derived from the baseline and fully adjusted models (Models 2 and 4, respectively); a table of significant pairwise comparisons between these predicted probabilities is presented in online appendix, Table S3. Consistent with findings from the logistic regression models, cisgender women have a significant and slightly higher predicted probability of reporting poor health than cisgender men (0.181 in both models). Transgender men have a significant health disadvantage compared with cisgender men in the baseline model (0.233), but this pairwise comparison is not significant after I adjust for demographic, socioeconomic, and behavioral factors in Model 4 (0.177), consistent with findings from the logistic regression estimates. Gender-nonconforming respondents have a significantly higher probability of reporting poor health compared with cisgender men, cisgender women, and transgender women in both models (0.303 in the baseline model and 0.278 in the fully adjusted model). However, gender-nonconforming respondents’ probabilities of poor health do not differ significantly compared with transgender men in the baseline or fully adjusted models.

Fig. 2
figure 2

Predicted probabilities of reporting poor health by gender identity and misclassification in BRFSS 2014–2016

Full size image

Although Hypothesis 1, which asserts that transgender respondents will have an overall health disadvantage compared with cisgender respondents, is confirmed by the logistic regression results, the predicted probabilities reveal serious heterogeneity within this group: transgender women do not have a significant health disadvantage compared with cisgender men in the predicted probabilities from either model. Gender-nonconforming respondents are the only noncisgender group with a consistently significant health disadvantage compared with both cisgender men and cisgender women. Transgender men have a significant disadvantage only in comparison with cisgender men in the baseline model. Hypothesis 2, testing the influence of being assigned female or male at birth, posits that cisgender women will have higher odds of poor self-rated health than cisgender men and that transgender men will have higher probabilities of poor self-rated health than transgender women. Cisgender women’s predicted probabilities of poor self-rated health are only slightly higher than those of cisgender men, and transgender men’s predicted probabilities of poor self-rated health are higher than cisgender men’s only in the baseline model and are never higher than those of transgender women in a significant pairwise comparison. Given the difference between cisgender women’s predicted health disadvantage in comparison with cisgender men and the lack of a discernable disadvantage between transgender men and women, these predicted probabilities suggest that being assigned a woman at birth does not necessarily predict a health disadvantage among transgender men.

Hypothesis 3, which posits that individuals who identify as women will have a health disadvantage compared with individuals who identify as men, does not have much evidence in its favor, either: I do not find significant differences in predicted probabilities of poor self-rated health between transgender men and transgender women, nor between cisgender men and cisgender women. I do find support for Hypothesis 4, which posits that gender-nonconforming respondents will have the highest health disadvantage of any group. Gender-nonconforming respondents had high probabilities of reporting poor health in both models (0.303 in the baseline model, 0.278 in the fully adjusted model) and significant contrasts in pairwise comparisons with cisgender men, cisgender women, and transgender women. Among all gender identity groups, predicted probabilities of reporting poor health are lower after adjusting for demographic, socioeconomic, and behavioral factors, except among cisgender men, whose probabilities of reporting poor health are slightly higher.

Conclusion

Through an intercategorical approach adapted from past research (Gorman et al. 2015), this study tests whether different configurations of sex categories (such as male or female assignment at birth) in conjunction with gender identities (such as man or woman, and whether one is transgender, cisgender, or gender-nonconforming) correspond to significant health differences that can be detected at the population level. Instead of testing for one basic association between cross-gender identification and self-rated health, grouping all transgender and gender-nonconforming identities together, I evaluate whether patterns of disadvantage are different for transgender women, transgender men, and gender-nonconforming individuals.

My findings suggest that gender-nonconforming individuals face significantly higher odds of reporting poor health compared with cisgender men and that transgender men and cisgender women also face some forms of this disadvantage. The estimates of reporting poor health among transgender men are sensitive to adjustments for socioeconomic factors, suggesting that transgender men’s health disparities may be driven by social exclusion and socioeconomic marginalization. Cisgender women have a persistent self-rated health disadvantage compared with cisgender men in all models, but transgender women’s odds of reporting poor health are not higher than those of transgender men. In fact, transgender women’s probabilities of poor health do not differ significantly from those of any other group, with the exception of a marked advantage in comparison with gender-nonconforming respondents in both baseline and fully adjusted models. I do not find evidence of a health disadvantage explicitly based on identifying as a woman, which cisgender women and transgender women share in common; however, in the fully adjusted model, compared with all groups, transgender women have the lowest odds of reporting poor health.

The persistence of the health disadvantage among gender-nonconforming individuals presents strong evidence of a significant association between identifying outside a binary gender identity (man or woman) and a higher risk of reporting poor health, at least among gender-nonconforming individuals who identify as transgender. This association between gender nonconformity and higher odds of reporting poor health persists even after I adjust for demographic and socioeconomic factors, marital status, and rates of ever having smoked: this group remains more than twice as likely as cisgender men to report poor health in all regression estimates and almost twice as likely in the fully adjusted predicted probabilities. Past research on the unique social position of gender-nonconforming individuals emphasizes the challenges of navigating social and clinical spaces in a cultural context in which belonging to the categories of “man” or “woman” is often treated as an essential requirement and in which instances of gender misclassification often cause stress or violent victimization for such individuals (Connell 2009; Harrison et al. 2012; James et al. 2016; Lombardi 2009; Schilt and Westbrook 2009). Gender-nonconforming respondents’ pronounced overall health disadvantage in comparison with all other gender identity groups underscores the need for the BRFSS and other health surveys to modify their data collection methods to be able to identify the sex respondents were assigned at birth. This is essential to understanding transgender and gender-nonconforming health at the population level.

The BRFSS sheds new light on questions related to the health of transgender and gender-nonconforming respondents, but several key limitations are relevant to this study. Because the data are cross-sectional, and social changes in gender identity can happen at any point in life, it is not possible to make causal claims about the relationship of gender identity to self-rated health, running counter to this study’s focus on establishing a predictive relationship. Furthermore, samples drawing from 31 states and one U.S. territory allow for larger generalizations about the U.S. transgender and gender-nonconforming populations than previously possible, but addition of the 19 remaining states, the District of Columbia, and four other U.S. territories, as well as ensuring a lower percentage of missing observations in future survey administrations, would inspire more confidence in the national representability of the sample. In addition to a more complete sample of U.S. states, a larger overall sample size would be advantageous to future work. The present sample size also does not permit an analysis of meaningful interactions between different categories of race/ethnicity and gender identity, even though it is crucial to approach gender identity with an intersectional lens that accounts for race/ethnicity (Nuttbrock et al. 2009; Sevelius 2013).

Another limitation of these data is that between and beyond the poles of identifying as transgender, gender-nonconforming, or cisgender lie so many ways of identifying and living with gender that cannot be fully encompassed by these categories. There is also reason to believe that among respondents who do not identify as cisgender, some might not identify themselves as transgender or gender-nonconforming to a survey interviewer (Westbrook and Saperstein 2015). Furthermore, as mentioned previously, the BRFSS methodology used to ascertain respondent sex has serious shortcomings (Riley et al. 2017). Future research could deepen what we know about gender identities and life processes by using methods that explicitly ask respondents about the sex they were assigned at birth as well as how they presently identify their gender (Reisner et al. 2015a, 2015b). However, this study pools samples from three years of data collection, consistent with the recommendations of the Gender Identity in U.S. Surveillance (GenIUSS) Group for mitigating bias from random and nonrandom error (Reisner et al. 2015a, 2015b).

To conclude, this article lays the groundwork for research at the population level that can distinguish between transgender men, transgender women, and gender-nonconforming individuals. I compare all three groups to the cisgender population, just as more research is starting to operationalize distinct forms of sexual orientation, race/ethnicity, and class as important social dimensions. This inclusion of expanded gender identity classification complicates what population health scholars are referring to when talking about “men” and “women.” I find that gender-nonconforming identity is associated with a marked health disadvantage compared with any other gender identity group. This highlights the need for future data collection efforts to collect sex assigned at birth separately from respondents’ gender identity in order to further differentiate the roles of sex and gender in shaping health patterns. Furthermore, the findings on transgender men’s health disadvantages suggest a need for more detailed research on their socioeconomic and social marginalization as well as its relationship to family composition. This study produces findings that appear to contradict qualitative and convenience sample–based studies that identify significant patterns of social and health disadvantages among transgender women (Brennan et al. 2012; Clements-Nolle et al. 2001; James et al. 2016; Schilt 2010). However, the lack of a discernable disadvantage in overall health does not mean that transgender women do not experience other significant disadvantages at the population level that could be identified in future work through other health-related measurements, such as HIV status and ability to access adequate health care.

Overall, these findings complicate facile claims about the relationships among the social norms that govern male or female assignment at birth, identifying as a man or woman, and identifying outside of these identities or experiences. As more large-scale social survey research captures these gender identities, future studies ought to consider the impact of gender-nonconforming identity in shaping overall health because it is apparent that this subpopulation experiences significant disadvantages. Future work should also more closely examine the roles of gender assignment at birth versus gender identity and delve more deeply into how these relationships change when they intersect with demographic, socioeconomic, and behavioral factors. Major health surveys, such as the BRFSS, need to make serious changes to their data collection methodologies in order for this future work to be possible, but the evidence available from existing data already indicates that transgender and gender-nonconforming populations face significant health disadvantages that require greater attention.