Exploring Social and Structural Determinants of Substance Use Among Sexual and Gender Minority Adults in the United States

Abstract

Purpose

To identify key social and structural factors associated with hazardous alcohol and problematic drug use among sexual and gender minority (SGM) adults.

Methods

We conducted a cross-sectional secondary analysis of 2016 to 2019 data from the Generations Study, a probability sample of 1518 SGM adults. Hazardous alcohol and problematic drug use were measured using the Alcohol Use Disorder Identification Test-Consumption (AUDIT-C) and Drug Use Disorder Identification Test (DUDIT). Candidate predictors included sociodemographic characteristics, minority stressors, adverse childhood experiences, health indicators, and social support. Interpretable machine-learning models with SHapley Additive exPlanations characterized multilevel correlates of 3-level risk categories.

Results

Hazardous alcohol use was linked to past smoking, lower social support, cancer history, housing instability, and identity-related stressors. Problematic drug use was linked to conversion-therapy exposure, housing discrimination, legal encounters, and mental-health indicators; prediction was more accurate for drug- than alcohol-use risk.

Conclusion

Structural vulnerability, discrimination, and identity-related stressors emerged as intervention targets to reduce substance-use disparities in SGM adults.

Plain Language Summary Title

What Social and Life Factors Affect Substance Use in Sexual Gender Minority Adults in the United States, Using Advanced Machine Learning Strategy to Explore

Plain Language Summary

This study looked at what social and life experiences put LGBTQ+ adults in the United States at higher risk for alcohol and drug problems. We analyzed survey data from the Generations Study 2016 to 2019 which included 1518 sexual and gender minority adults (eg, gay, lesbian, transgender, queer, and bisexual).

Alcohol use was measured by how often and how much people drank, while drug use was measured by the frequency and impact of nonalcohol substance use. We also considered many possible influences, such as age, income, childhood experiences, discrimination, health, and support from family and friends.

For alcohol use, the most important factors were past smoking, low social support, cancer, unstable housing, and identity-related stress. For drug use, key factors included exposure to conversion therapy, housing discrimination, legal problems, and mental health concerns such as self-harm. Our models were better at predicting drug use than alcohol use.

These findings show that alcohol and drug use among LGBTQ+ people is shaped by personal behavior and also by stress, discrimination, and unequal access to stable housing, health, and support. Policies and programs to reduce substance use need to go beyond individual choices and address these broader social and structural conditions. Using transparent machine learning methods, we were able to identify complex patterns that point to where interventions and resources are most needed.

Keywords

substance use minority stress sexual and gender minorities health disparities structural determinants

Introduction

Sexual and gender minority (SGM) individuals, including those who identify as lesbian, gay, bisexual, transgender, queer, nonbinary, and other nonheteronormative identities, experience disproportionately high rates of substance use, including hazardous alcohol and drug consumption, relative to cisgender and heterosexual populations.^1–3 These disparities are not solely attributable to individual-level behaviors or choices but are rooted in broader systems of oppression, discrimination, and marginalization that operate across multiple socioecological levels.^4–6

The Minority Stress Model^7–9 posits that stigma-related stressors (eg, discrimination, victimization, anticipated rejection, and internalized stigma) accumulate over time, contributing to adverse mental health and coping behaviors, including substance use. Intersectionality Theory^10,11 adds that multiple marginalized positions (eg, race, socioeconomic status, gender identity) intersect to produce compounded disadvantage.^11–14 The Social Determinants of Health (SoDH) framework^15–17 further emphasizes how structural conditions (eg, housing instability, discrimination, and criminal-legal exposure) are associated with risk environments beyond individual agency. Together, these lenses underscore how structural inequities and identity-based stressors interact to impact substance-use risk.

However, most existing research has relied on statistical methods that assume linear relationships, independence among variables, and prespecified hypotheses.¹⁸ These methods may fail to capture the complexity of the factors influencing substance use, particularly in populations facing layered social vulnerabilities. There remains a critical need for analytic approaches that can accommodate the high dimensionality and nonlinearity inherent in social and behavioral data.^19,20 Alternatively, machine learning (ML) algorithms can model complex interactions among many variables without relying on prior assumptions about their relationships. Additionally, recent advances in interpretable ML, such as SHapley Additive exPlanations (SHAP),²¹ allow researchers to retain insight into variable importance and effect directionality, addressing longstanding criticisms of ML's “black box” nature.^19,22–24 In the context of SGM health research, ML may uncover previously underrecognized combinations of risk and protective factors spanning individual, interpersonal, and structural domains that influence substance use.

Therefore, we applied interpretable ML techniques that were guided by the Minority Stress Model and the SoDH Framework to: (1) prioritize multilevel correlates of hazardous alcohol use and problematic drug use among SGM adults and (2) translate these findings to actionable screening, clinical workflow, and policy targets. This approach preserves theoretical interpretability (via SHAP value) while accommodating high-dimensional, nonlinear patterns to bridge explanatory theory and decision support. Because alcohol and drug use can elevate HIV risk through decreased PrEP uptake/adherence, condomless sex, and cooccurring mental health burden, identifying upstream social/structural drivers in SGM communities also informs HIV prevention strategies.

Methods

Study Design and Participants

This study employed a cross-sectional secondary data analysis of baseline data (collected 2016-2019) from the Generations Study, which was designed to examine the health, stress, and well-being of SGM adults.²⁵ The study employed a national probability sample drawn from the Gallup Daily Tracking Survey, which included respondents aged 18 and older from all 50 U.S. states and the District of Columbia. Participants were screened for eligibility using a 2-phase process. In the first phase, individuals were identified via a self-reported question on sexual orientation and gender identity. Then, eligible participants were invited to complete a self-administered survey, either online or by mail. The final dataset included responses from 1518 individuals, with an oversample of Black and Latino participants to ensure adequate representation of these groups. Recruitment procedures ensured diversity in race/ethnicity, geography, and socioeconomic status. Participants provided informed consent and received compensation for their time.²⁵ This is a cross-sectional secondary analysis that is exploratory in nature and aims to develop and evaluate predictive models for substance-use risk stratification (alcohol and drug use). Reporting follows the STROBE guidelines (Supplemental Table S1).

Outcomes

The primary outcome variables in this study were alcohol use and drug use, assessed using validated instruments. Alcohol use was measured with the Alcohol Use Disorder Identification Test-Consumption (AUDIT-C), a widely used 3-item scale that evaluates hazardous drinking behaviors and potential alcohol-use disorders.^26,27 Scores ranged from 0 to 12, with higher scores indicating greater levels of risky alcohol consumption (Cronbach's alpha = .790). For the AUDIT scale, scores of 0 to 3 were classified as low risk, 4 to 6 as moderate risk, and ≥7 as high risk. Drug use was assessed using the Drug Use Disorder Identification Test (DUDIT), an 11-item scale designed to identify individuals with drug-related problems.^28,29 DUDIT scores ranged from 0 to 44, with thresholds distinguishing substance abuse from dependence (Cronbach's alpha = .856). For the DUDIT scale, scores of 1 to 10 were classified as low risk, 11 to 15 as moderate risk, and ≥16 as high risk. For both alcohol and drug use, we interpret the 3-level categories (low, moderate, high risk) as reflecting increasing severity of hazardous alcohol use (AUDIT-C) and problematic drug use (DUDIT).

Potential Variables Selected for Model Training

While the outcome variables were the focus of the analysis, variables were selected from a comprehensive dataset spanning multiple domains, including stressors, health outcomes, and healthcare utilization. We selected a series of potential variables informed by the Minority Stress Model, Intersectionality Theory, and the SDoH framework. These variables reflected multiple domains relevant to SGM health, including: (1) sociodemographic factors: age, gender identity, race/ethnicity, education, income; (2) minority stress indicators: enacted discrimination (lifetime and past-year events), anticipated rejection/expectations of rejection, internalized stigma, and identity concealment/outness milestones; we also included conversion-therapy exposure and housing/healthcare discrimination (binary indicators); (3) adverse life experiences: adverse childhood experiences (ACE), family rejection, housing instability, incarceration history; (4) mental and physical health: depressive symptoms, chronic illness (eg, kidney disease), history of self-injury, access to healthcare; and (5) social support and community connection: support from family or friends, identity affirmation (total score; higher = more support), connection to LGBTQ+ communities. We excluded any variables that directly overlapped with items used in the AUDIT-C or DUDIT to avoid artificially inflating associations. To prevent information leakage, we excluded items that directly overlapped with AUDIT-C or DUDIT content. Variables were not prespecified; instead, we used feature-selection algorithms (described below) to identify the most influential predictors of alcohol- and drug-use risk.^30,31

Data Preparation

Data preprocessing included excluding variables with more than 20% missing data. Remaining missing values (<20%) were imputed using k-nearest neighbor imputation to preserve sample size while maintaining plausible estimates. This strategy balanced data completeness and model reliability by removing variables with excessive missingness, thereby minimizing bias and improving interpretability.^32,33 Questions directly related to alcohol or drug use were excluded from the variables to prevent information leakage. Noninformative columns (eg, study IDs) were removed. Highly correlated variables were screened for multicollinearity, and 1 variable from each correlated pair (Pearson r > .80) was dropped. Categorical variables were binarized, and continuous variables were standardized for comparability.

Statistical Analysis

All analyses were conducted in Python 3.9. The dataset was randomly partitioned into training (80%) and testing (20%) sets, stratified by outcome class (low, moderate, high risk). Within the training set, a 20% validation split was used for hyperparameter tuning and early stopping. All preprocessing procedures (imputation, encoding, scaling, and correlation filtering) were fit on the training data and then applied to validation and test sets to prevent information leakage.

To identify stable and robust predictors, we implemented a feature-selection strategy based on a weighted voting scheme that aggregates feature importance from Elastic Net (EN), Random Forest (RF), and XGBoost (XGB) models. EN provides a linear, sparse representation that handles multicollinearity, whereas RF and XGB capture nonlinearities and higher-order interactions.³⁴ For each model, feature importance scores were min–max normalized to the 0 to 1 range and combined as I_combo = w_enI_en + w_rfI_rf ⁺ w_xgbI_xgb (where I_combo, I_en,I_rf, I_xgb are the feature-importance scores from each model; and w_en, w_rf, w_xgb are the nonnegative weights assigned to each model that sum to 1). We evaluated a grid of weight combinations (0.0-0.9 in 0.1 increments) and, for each combination, selected the top 30 features (k = 30), chosen a priori based on the elbow observed in the cumulative importance/predictive plateau. Features were then ranked by selection frequency across weightings and stratified cross-validation folds, and the final feature set was chosen to maximize stability.

Using the selected feature set, we trained 4 multiclass classifiers—XGB, RF, support vector machine, and a shallow feed-forward neural network (NN)—to predict 3-level risk categories (low, moderate, high) for both alcohol (AUDIT-C) and drug (DUDIT) outcomes. For analyses of the full Wave 1 sample (original plus enhancement), we incorporated the Wave 1 probability weight (weight_full) provided by the Generations Study. After preprocessing (imputation, encoding, scaling, correlation filtering), we applied the stability-oriented feature selection procedure described above (weighted voting across EN, RF, and XGB importance scores) and then trained all models using stratified train/validation/test splits to prevent information leakage. The NN comprised 3 fully connected layers with rectified linear unit activations and a SoftMax output layer; to reduce overfitting, we used dropout (rate = 0.30) on hidden layers and early stopping based on validation loss (Adam optimizer, learning rate 1 × 10⁻³, batch size 64, up to 100 epochs, patience = 10). Model interpretability was assessed using SHAP values, which quantify the direction and relative contribution of each feature to predicted risk. For descriptive comparisons of demographic and key clinical characteristics across alcohol- and drug-use risk categories, we used analysis of variance with appropriate post-hoc tests.

Results

Table 1 shows the distribution of alcohol and drug use across various sociodemographic factors among the study participants. Age was significantly associated with both alcohol and drug use, with individuals in the high-risk drug use group being notably younger (mean age = 33.28 years, standard deviation [SD] = 14.06) compared to those in the low-risk group (mean age = 38.94 years, SD = 15.83, η² = 0.01), suggesting a small effect. Gender identity had a minor impact on alcohol use (Cramer's V = 0.05) and a slightly larger effect on drug use (Cramer's V = 0.06), with men demonstrating higher rates of drug use compared to women and nonbinary individuals. Similarly, racial and ethnic disparities were evident in substance use patterns. White participants had the highest representation in the high-risk alcohol use group (Cramer's V = 0.03), while Black/African American participants had the lowest representation in the high-risk drug use group (Cramer's V = 0.07). However, effect sizes suggest small to moderate associations. Educational attainment was also associated with substance use, with individuals holding a bachelor's degree or higher being more likely to be in the low-risk categories for both alcohol (Cramer's V = 0.09) and drug use (Cramer's V = 0.09). Employment status had a moderate effect on alcohol use (Cramer's V = 0.13) and a more negligible effect on drug use (Cramer's V = 0.08), as unemployed individuals were overrepresented in high-risk groups for both substances. Census poverty demonstrated a minor impact on drug use (Cramer's V = 0.08), with individuals living in poverty exhibiting higher rates of problematic and hazardous drug use. Self-reported general health status was also linked to substance use, with those reporting excellent or very good health being more likely to be in the low-risk alcohol (Cramer's V = 0.08) and drug (Cramer's V = 0.09) use categories (Table 1).

Table 1.

Demographic Information for Participants.

		Alcohol use				Drug use
Variable	Total	Low	Medium	High	Effect size* (p-value)	Low	Medium	High	Effect size
Age (mean, sd)	1502	37.74 (sd = 15.76)	35.00 (sd = 14.29)	36.93 (sd = 13.84)	0.01 (.01)	38.94 (sd = 15.83)	37.42 (sd = 14.71)	33.28 (sd = 14.06)	0.01 (<.01)
Sex assigned at birth	1488 (100)	476 (31.99)	558 (37.5)	454 (30.51)		32 (2.15)	1103 (74.13)	353 (23.72)
Male (n, %)	688 (46.24)	208 (30.23)	250 (36.34)	230 (33.43)	0.06 (.07)	14 (2.03)	509 (73.98)	165 (23.98)	0.01 (.94)
Female (n, %)	800 (53.76)	268 (33.50)	308 (38.50)	224 (28.00)	0.06 (.07)	18 (2.25)	594 (74.25)	188 (23.50)	0.01 (.94)
Gender identity	1495 (100)	479 (32.04)	560 (37.46)	456 (30.50)		32 (2.14)	1109 (74.18)	354 (23.68)
Man (n, %)	662 (44.28)	194 (29.31)	245 (37.00)	223 (33.69)	0.05 (.11)	14 (2.11)	494 (74.62)	154 (23.26)	0.06 (.02)
Women (n, %)	740 (49.50)	250 (33.78)	283 (38.24)	207 (27.97)		15 (2.03)	560 (75.68)	165 (22.30)
Nonbinary/others (n, %)	93 (6.22)	35 (37.63)	32 (34.41)	26 (27.96)		3 (3.23)	55 (59.14)	35 (37.63)
Race/ethnicity	1502 (100)	480 (31.96)	564 (37.56)	458 (30.48)		32 (2.11)	1113 (74.11)	357 (23.77)
White (n, %)	926 (61.65)	296 (31.97)	340 (36.72)	290 (31.32)	0.03 (.74)	24 (2.59)	703 (75.92)	199 (21.49)	0.07 (.03)
Black/African American (n, %)	179 (11.92)	63 (35.20)	67 (37.43)	49 (27.37)		0 (0.00)	135 (75.42)	44 (24.58)
Latino (n, %)	156 (10.39)	52 (33.33)	57 (36.54)	47 (30.13)		2 (1.28)	113 (72.44)	41 (26.28)
Others (n, %)	241 (16.04)	69 (28.63)	100 (41.49)	72 (29.88)		6 (2.49)	162 (67.22)	73 (30.29)
Sexual identity	1502 (100)	480 (31.95)	564 (37.55)	458 (30.49)		32 (2.11)	1113 (74.11)	357 (23.77)
Lesbian (%)	301 (20.04)	119 (39.53)	101 (33.55)	81 (26.91)	0.06 (.09)	11 (3.65)	242 (80.40)	48 (15.95)	0.10 (<.01)
Gay (%)	534 (35.55)	157 (29.40)	208 (38.95)	169 (31.65)		12 (2.25)	410 (76.78)	112 (20.97)
Bisexual (%)	487 (32.42)	152 (31.21)	188 (38.60)	147 (30.18)		6 (1.23)	343 (70.43)	138 (28.34)
Others (%)	180 (11.98)	52 (28.89)	67 (37.22)	61 (33.89)		3 (1.67)	118 (65.56)	59 (32.78)
Education	1502 (100)	480 (31.96)	564 (37.55)	458 (30.49)		32 (2.13)	1113 (74.11)	357 (23.76)
High school or less (%)	306 (20.37)	118 (38.56)	121 (39.54)	67 (21.90)	0.09 (<.01)	6 (1.96)	205 (66.99)	95 (31.05)	0.09 (<.01)
Some college (%)	488 (32.49)	168 (34.43)	184 (37.70)	136 (27.87)		6 (1.23)	349 (71.52)	133 (27.25)
Bachelor's degree or higher (%)	708 (47.14)	194 (27.40)	259 (36.58)	255 (36.02)		20 (2.82)	559 (78.95)	129 (18.22)
Employment status	1438 (100)	460 (31.99)	534 (37.13)	444 (30.87)		32 (2.23)	1068 (74.26)	338 (23.51)
Employed full-time (%)	867 (60.29)	228 (26.30)	327 (37.72)	312 (35.99)	0.13 (<.01)	21 (2.42)	666 (76.82)	180 (20.76)	0.08 (.002)
Employed part-time (%)	259 (18.01)	87 (33.59)	108 (41.70)	64 (24.71)		5 (1.93)	168 (64.86)	86 (33.20)
Unemployed (%)	312 (21.70)	145 (46.47)	99 (31.73)	68 (21.79)		6 (1.92)	234 (75.00)	72 (23.08)
Census poverty	1482 (100)	472 (31.85)	556 (37.52)	454 (30.63)		32 (2.16)	1096 (73.95)	354 (23.89)
In poverty (%)	208 (14.04)	67 (32.21)	85 (40.87)	56 (26.92)	0.04 (.40)	3 (1.44)	137 (65.87)	68 (32.69)	0.08 (.005)
Not in poverty (%)	1274 (85.96)	405 (31.79)	471 (36.97)	398 (31.24)	0.04 (.40)	29 (2.28)	959 (75.27)	286 (22.45)	0.08 (.005)
General health status	1491 (100)	476 (31.93)	560 (37.56)	455 (30.51)		31 (2.08)	1106 (74.18)	354 (23.74)
Excellent/Very good (%)	753 (50.50)	218 (28.95)	275 (36.52)	260 (34.53)	0.08 (<.01)	17 (2.26)	598 (79.42)	138 (18.33)	0.09 (<.01)
Good (%)	478 (32.06)	154 (32.22)	189 (39.54)	135 (28.24)		11 (2.30)	323 (67.57)	144 (30.13)
Fair/Poor (%)	260 (17.44)	104 (40.00)	96 (36.92)	60 (23.08)		3 (1.15)	185 (71.15)	72 (27.69)
ACE^& (mean, sd)	1502	3.24 (sd = 2.12)	3.20 (sd = 2.19)	3.00 (sd = 2.12)	0.002 (.21)	3.28 (sd = 1.77)	2.95 (sd = 2.10)	3.78 (sd = 2.19)	0.027 (<.01)
ACE^&	1449 (100)	459 (31.67)	545 (37.61)	445 (30.71)		32 (2.20)	1075 (74.19)	342 (23.61)
Yes (%)	892 (61.56)	270 (30.27)	331 (37.11)	291 (32.62)	0.05 (.11)	19 (2.13)	688 (77.13)	185 (20.74)	0.09 (<.01)
No (%)	557 (38.44)	189 (33.93)	214 (38.42)	154 (27.65)	0.05 (.11)	13 (2.33)	387 (69.48)	157 (28.19)	0.09 (<.01)

Notes. ^&ACE, adverse childhood experience.

*Effect sizes are Cramer's V for chi-squares test, and eta-square for ANOVA.

As shown in Supplemental Tables 2a–2b, post-hoc analyses indicated that hazardous alcohol users and problematic drug users were significantly younger and reported higher ACE scores compared with low- and medium-risk users (p < .01). For categorical variables, pairwise chi-square tests (ie, Holm–Bonferroni adjusted) confirmed that differences in alcohol and drug use risk were most pronounced across education, employment, general health status, and sexual identity categories. In contrast, gender and race/ethnicity differences were minor. Participants with higher education, full-time employment, and very good or excellent health were consistently more likely to fall into the low-risk groups for both alcohol and drug use. In contrast, those reporting economic hardship or poorer health were disproportionately represented among higher-risk users.

Tables 2a and 2b present feature importance rankings and performance metrics from ML models examining alcohol and drug use among SGMs, identifying the most influential features across models. Across all models, social, behavioral, and mental health variables emerged as dominant variables of substance use risk, although the relative influence of specific features varied by model and substance type. For alcohol use, model performance was moderate, with F1 scores ranging from 0.31 to 0.47 and accuracy ranging from 0.37 to 0.44. The XGB model achieved the best balance (F1 = 0.47), while the NN demonstrated consistent performance across both alcohol and drug use tasks (F1 = 0.42 for AUDIT). Lifetime smoking history emerged as the most influential feature in the XGB model, followed by social and identity-related factors such as limited social support, higher frequency of moving, and milestones in sexual identity development (eg, coming-out age to straight friends and age of same-sex relationship). Financial and housing stability indicators, including household income and living arrangements (renting vs owning), were also among the top variables, highlighting protective roles of economic security and residential stability. In addition, recent poor mental or physical health and suicidal ideation contributed substantially to elevated alcohol use risk, underscoring the intersection of psychosocial stress and health vulnerability among SGMs (Table 2a). For drug use, models demonstrated stronger predictive performance, with F1-scores ranging from 0.59 to 0.68 and accuracy values between 0.71 and 0.75, indicating that drug use risk patterns were more distinct and more straightforward to classify than alcohol use patterns. Across models, the most influential features included exposure to conversion therapy, lower perceived social support, and experiences of victimization or discrimination (eg, religious or employment-related). Mental health factors such as poor mental health days, suicidal thoughts, and social well-being ranked prominently, alongside life-course and relational variables, including age of coming out, same-sex relationship age, and family discovery of sexual identity. Health-related factors, such as kidney disease, ACE scores, and STI testing frequency, also contributed to predictive accuracy. The NN model achieved the most stable and generalized performance across both AUDIT and DUDIT outcomes (F1 = 0.42 and 0.68, respectively), reflecting its capacity to capture complex, shared feature representations within the multitask learning framework (Table 2b). Among all classifiers, the NN achieved the most balanced performance across low, moderate, and high-risk categories, with F1-scores that were relatively similar across classes. Supplemental Figures 1-4 displayed SHAP summary plots for the final models, where feature values on the X-axis were standardized (z-scored) rather than raw units (eg, years of age or income in dollars), so that higher values represented relatively higher levels of each characteristic within the sample.

Table 2a.

Top 10 Features of the Prediction Model for Alcohol Use Among the LGB Population.

XGBoost				Support vector machine				Random Forest				Neural network
Low risk		High risk		Low risk		High risk		Low risk		High risk		Low risk		High risk
Key features	Importance	Key features	Importance	Key features	Importance	Key features	Importance	Key features	Importance	Key features	Importance	Key features	Importance	Key features	Importance
Move frequency	0.186250	Smoke 100 cigarettes for the entire life	0.329291	Distance from the LGB health clinic	0.010037	Same-sex relationship age	0.016619	Move frequency	0.016923	Smoke 100 cigarettes for the entire life	0.021006	Gender identify_femme	0.088579	Smoke 100 cigarettes for the entire life	0.134475
Attracted to men	0.150465	People living on an income	0.242859	Coming-out age_ straight friend	0.006869	Distance from the LGB health clinic	0.011007	Smoke 100 cigarettes entire life	0.013146	People living on an income	0.016462	Living arrangement_rent	0.080334	Poor mental health in the past 30 days	0.068194
Selfless help	0.147621	Social stagnation	0.147543	Same-sex first sex	0.005948	Same-sex first sex	0.010695	Attracted to men	0.012158	Distance from LGBT health clinic	0.012832	Attracted to men	0.079049	People living on an income	0.064853
Racial identity	0.134818	Same-sex relationship age	0.140350	Same-sex relationship age	0.005731	LGB Family discovery age	0.002285	Living arrangement_rent	0.011875	Living arrangement_rent	0.010736	Social support	0.077938	Social stagnation	0.064268
Social support	0.119542	Same-sex first sex	0.131043	Age of thinking about suicide	0.005658	Coming-out age_straight friend	0.001004	Sexual partner number	0.010286	Same-sex relationship age	0.010605	Poor mental health in the past 30 days	0.077216	Social support	0.063024
Gender identify_female	0.114432	Distance from the LGB health clinic	0.127436	LGB family discovery age	0.002707	Present religion	0.000933	Selfless help	0.009277	Religious attendance	0.010309	Religious attendance	0.064448	Education	0.061247
Coming-out age_straight friend	0.109516	Poor health for the past 30 days	0.116649	Coming-out age family	0.002039	Age of thinking about suicide	0.000872	Outness at work	0.008958	Move frequency	0.009276	Total annual household income	0.062662	Special equipment needed	0.056511
Distance from the LGB health clinic	0.105846	Present religion	0.110592	Realize the LGB age	0.001908	Coming-out age family	0.000831	Religious attendance	0.008784	Present religion	0.009195	Suicide thoughts	0.052262	Living arrangement_rent	0.054266
Coming-out age family	0.100225	Feel lonely	0.108230	Same-sex attraction age	0.001032	Poor physical health in the past 30 days	0.000811	People living on an income	0.008510	Education	0.008730	Racial identity	0.047750	ACE	0.053658
Religious attendance	0.095825	Religious attendance	0.106251	Suicide attempt age	0.000678	Realize the LGB age	0.000502	Job changes last year	0.008369	Annual personal income	0.008572	Age of thinking about suicide	0.042439	Suicide thoughts	0.047787
Accuracy	0.44			0.37				0.41				0.42
Precision	0.47			0.44				0.37				0.42
Recall	0.47			0.43				0.37				0.42
F1-score	0.47			0.43				0.37				0.42

Notes. ACE: adverse childhood experiences.

Table 2b.

Top 10 Features of the Prediction Model for Drug Use Among the LGB Population.

XGBoost				Support vector machine				Random Forest				Neural network
Low risk		High risk		Low risk		High risk		Low risk		High risk		Low risk		High risk
Key features	Importance	Key features	Importance	Key features	Importance	Key features	Importance	Key features	Importance	Key features	Importance	Key features	Importance	Key features	Importance
Treatment to change sexual orientation	0.233148	Smoke 100 cigarettes for the entire life	0.382677	Same-sex relationship age	0.000109	Distance from the LGB health clinic	0.000931	Coming-out age_straight friend	0.012827	Smoke 100 cigarettes for the entire life	0.031517	COHORT	0.046410	Smoke 100 cigarettes for the entire life	0.063639
Coming-out age family	0.224224	Distance from the LGBT health clinic	0.178124	Distance from the LGB health clinic	0.000088	Age of thinking about suicide	0.000560	Same-sex attraction age	0.011978	Poor health for the past 30 days	0.015695	Proud of LGBT	0.040482	Victimization frequency	0.050413
Social support	0.198772	Poor health for the past 30 days	0.175398	Coming-out age_straight friend	0.000056	Same-sex relationship age	0.000468	ACE	0.011291	STI test frequency	0.015561	Same-sex attraction age	0.026267	STI test frequency	0.049539
Proud of LGB	0.198100	Same-sex first sex	0.168829	Age of thinking about suicide	0.000045	Poor mental health past 30 days	0.000462	Social support	0.010532	Poor mental health in the past 30 days	0.013796	Past sexual activity_transgender	0.026258	Poor mental health in the past 30 days	0.042227
Live with people who use drugs	0.192327	STI test frequency	0.150858	Same-sex first sex	0.000036	LGB Family discovery age	0.000310	Realize the LGB age	0.010504	ACE	0.011951	HIV worry frequency	0.022912	Insurance through parents	0.041101
Distance from the LGB health clinic	0.175701	Poor mental health in the past 30 days	0.142254	Coming-out age family	0.000035	Future social ladder	0.000272	Same-sex relationship age	0.010378	Same-sex first sex	0.011390	Have children	0.022611	Proud of LGBT	0.035636
ACE	0.165962	Victimization frequency	0.123746	Age of thinking about suicide	0.000031	Same-sex attraction age	0.000253	Coming-out age family	0.010061	Live with people who use drugs	0.011311	Military service	0.022448	Perceived social uselessness	0.034236
LGB family discovery age	0.145648	Clinic place: doctor's office or HMO	0.123610	LGB family discovery age	0.000027	Same-sex first sex	0.000215	Social well-being	0.010004	Distance from the LGB health clinic	0.010591	Kidney disease	0.022103	Social support	0.031892
Sexual orientation	0.144387	Employment equality perception	0.112912	Realize the LGB age	0.000025	STI test frequency	0.000204	Distance from the LGB health clinic	0.009914	Current social ladder	0.010253	Annual personal income	0.020471	Clinic place: doctor's office or HMO	0.030459
HIV test frequency	0.139424	Future social ladder	0.109508	Future social ladder	0.000018	Coming-out age_straight friend	0.000166	Smoke 100 cigarettes for the entire life	0.009230	Perceived social uselessness	0.009898	Perceived religious discrimination	0.019893	W1Q135C	0.027845
Accuracy	0.44			0.37				0.41				0.42
Precision	0.47			0.44				0.37				0.42
Recall	0.47			0.43				0.37				0.42
F1-Score	0.47			0.43				0.37				0.42

Notes. ACE, adverse childhood experiences.

Discussion

This study examined associations between social, psychological, and structural factors and alcohol and drug use among SGM adults using a theoretically grounded, data-driven approach. By applying a combination of ML techniques to a large, nationally representative dataset, we were able to surface both expected and underrecognized factors of substance use across 2 distinct domains: hazardous alcohol use and problematic drug use. While the predictive performance varied by outcomes, our findings reveal meaningful patterns that expand our understanding of substance use risk in SGM groups.

The alcohol use models highlighted factors that resonate with existing literature but also underscore the compounding influence of structural disadvantages and interpersonal stress.^35–38 Limited family or social support, particularly in the context of stigma or rejection, further increased the risk of hazardous alcohol use. Notably, housing instability, a widely recognized social determinant of health, was a prominent feature, suggesting that economic precarity and displacement may exacerbate vulnerability to alcohol misuse. The relevance of internalized stigma, identity concealment, and discomfort with specific identity labels suggests that intracommunity dynamics and identity development processes deserve greater attention in alcohol-focused interventions. On the other hand, the alcohol models showed modest discrimination relative to the drug models. This pattern is plausible given that alcohol use is more normative and context-dependent, shaped by social settings, availability/outlet density, and weekday/seasonal routines that are undermeasured in the present dataset.

In contrast, drug use models revealed a more overt pattern of structural and psychological harm. The strong role of conversion therapy exposure aligns with prior findings that such experiences are linked to increased psychological distress, trauma symptoms, and maladaptive coping.^39,40 Housing discrimination and criminal legal involvement further exemplify how institutionalized forms of stigma and exclusion contribute to elevated drug use risk. These results are consistent with the Minority Stress Model^4,7,8,41 and affirming the SoDH framework.^15,16,42 Both frameworks emphasize how material deprivation and systemic bias influence behavioral outcomes.

Our findings carry several implications for intervention and policy. The prioritized social/structural correlates suggest actionable interventions: (1) housing stability supports (eg, rental assistance, eviction protections) in jurisdictions with high housing cost burden; (2) healthcare antidiscrimination enforcement and staff training to reduce stigma in clinical and pharmacy settings; (3) statewide conversion-therapy bans and enforcement; (4) expansion of LGBTQ-affirming primary care and integrated behavioral health; (5) streamlined access to PrEP (same-day starts, confidential or on-site pharmacies, reduced prior authorization); and (6) legal aid and record-expungement programs where criminal-legal exposure emerged as salient.

Our study also highlights the value of ML as a complementary tool for public health equity research.^19,20,43,44 Although tree-based methods such as XGB and RF provided strong overall performance, the NN demonstrated the most balanced performance across risk categories. This balanced profile suggests that the NN was less dominated by the majority (low risk) class and retained reasonable discrimination for participants at moderate and high risk of hazardous alcohol use and problematic drug use. This is particularly important in the context of SGM substance use, where the high-risk substance use groups are relatively small but clinically critical, and it suggests that flexible, nonlinear models can improve risk stratification for both lower- and higher-risk SGM adults, rather than primarily optimizing prediction for the majority low-risk group. While the alcohol models showed only modest discrimination, the convergence of salient features across algorithms supports the robustness of the observed patterns. Using SHAP value, we retained transparency about how stressors and supports relate to substance-use classifications, clarifying the direction and relative contribution of multilevel correlates. Because the SHAP plots are based on standardized feature values, the patterns should be interpreted as relative increases or decreases in each predictor (eg, higher vs lower age or income) rather than specific raw units. These models are not intended for diagnosis or deployment; rather, they are hypothesis-generating and designed to inform intervention prioritization and policy strategy. Beyond reaffirming known correlations, interpretable ML highlighted nonlinearities, interactions, and threshold-like effects that conventional linear models may miss. For example, the cooccurrence of stigma and depressive symptoms, and unstable housing with poverty, was linked to disproportionate increases in risk in SHAP interaction plots, even when each factor showed only small bivariate associations. Together, these insights identify that cooccurring conditions that are most strongly linked to hazardous alcohol use or problematic drug use and point to concrete targets for coordinated screening and structural action.

Limitations

Several limitations warrant consideration. First, although we applied for the Generations Study survey weights, residual sampling, and nonresponse bias may persist and could affect generalizability. Second, key variables, including alcohol/drug use and stigmatization, were self-reported, and thus subject to social desirability and recall bias. Third, the cross-sectional design precludes causal inference; results should be interpreted as associations rather than effects. Fourth, bivariate effect sizes were generally small, consistent with complex, multilevel phenomena in which individual variables exhibit modest associations, but combinations of factors carry a stronger signal. Fifth, alcohol consumption may be especially influenced by unmeasured social and environmental norms, which could contribute to the comparatively lower discrimination of alcohol models in this study. Sixth, limited cell sizes for some subgroups (eg, nonbinary participants, Black transgender participants) constrained stratified analyses; future work should oversample underrepresented SGM subgroups to enable precision estimates.

Conclusion

A convergence of identity-based stressors, interpersonal exclusion, and structural inequities shapes high-risk substance use among SGMs.^4–6 Our study adds to the growing body of work that calls for comprehensive, trauma-informed, and equity-centered approaches to substance use prevention and care in SGM populations. By combining theory, empirical data, and modern analytic tools, we offer a framework for identifying those most at risk—and the structural levers needed for changes.

Supplemental Material

sj-docx-1-jia-10.1177_23259582251409855 - Supplemental material for Exploring Social and Structural Determinants of Substance Use Among Sexual and Gender Minority Adults in the United States

Supplemental material, sj-docx-1-jia-10.1177_23259582251409855 for Exploring Social and Structural Determinants of Substance Use Among Sexual and Gender Minority Adults in the United States by Chen Zhang, Yao Tang, Wonkyung Kniffen and Yu Liu in Journal of the International Association of Providers of AIDS Care (JIAPAC)

Supplemental Material

sj-docx-2-jia-10.1177_23259582251409855 - Supplemental material for Exploring Social and Structural Determinants of Substance Use Among Sexual and Gender Minority Adults in the United States

Supplemental material, sj-docx-2-jia-10.1177_23259582251409855 for Exploring Social and Structural Determinants of Substance Use Among Sexual and Gender Minority Adults in the United States by Chen Zhang, Yao Tang, Wonkyung Kniffen and Yu Liu in Journal of the International Association of Providers of AIDS Care (JIAPAC)

Supplemental Material

sj-docx-3-jia-10.1177_23259582251409855 - Supplemental material for Exploring Social and Structural Determinants of Substance Use Among Sexual and Gender Minority Adults in the United States

Supplemental material, sj-docx-3-jia-10.1177_23259582251409855 for Exploring Social and Structural Determinants of Substance Use Among Sexual and Gender Minority Adults in the United States by Chen Zhang, Yao Tang, Wonkyung Kniffen and Yu Liu in Journal of the International Association of Providers of AIDS Care (JIAPAC)

Supplemental Material

sj-pdf-4-jia-10.1177_23259582251409855 - Supplemental material for Exploring Social and Structural Determinants of Substance Use Among Sexual and Gender Minority Adults in the United States

Supplemental material, sj-pdf-4-jia-10.1177_23259582251409855 for Exploring Social and Structural Determinants of Substance Use Among Sexual and Gender Minority Adults in the United States by Chen Zhang, Yao Tang, Wonkyung Kniffen and Yu Liu in Journal of the International Association of Providers of AIDS Care (JIAPAC)

Footnotes

Acknowledgments

This publication is funded by the University of Rochester, the Rita and Alex Hillman Foundation, and the generosity of Leonard A. Lauder. The Rita and Alex Hillman Foundation (the “Foundation”) is one of the nation's leading philanthropies dedicated to advancing health equity and improving the health and healthcare of marginalized populations. The Foundation supports innovative, nursing-driven models of care that address the needs of communities experiencing inequity, discrimination, oppression, and indifference. The contents of this manuscript do not necessarily reflect the opinions of the funders.

ORCID iDs

Chen Zhang

Yu Liu

Ethical Approval and Written Informed Consent Statements

The University of Rochester Research Subjects Review Board reviewed this project (Study ID: STUDY00011070) and determined that it does not constitute human subjects research because it involves secondary analysis of a fully deidentified dataset (ICPSR 37166). As such, additional IRB approval and informed consent were not required for this analysis.

Authorship Contribution

CZ conceived and designed the study, supervised the analytic plan, and drafted the manuscript; YT conducted data management, coding, and ML analyses, and contributed to the interpretation of the findings; WK contributed to study design, interpretation of results in relation to mental health and clinical implications, and critically revised the manuscript for important intellectual content; YL contributed to conceptualization, provided methodological and statistical guidance, and substantially revised the manuscript for clarity and interpretation. All authors reviewed and approved the final version of the manuscript and agree to be accountable for all aspects of the work.

Funding

This study was sponsored by the Research Award of the University of Rochester and the Rita and Alex Hillman Foundation.

Declaration of Conflicting Interests

The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Data Availability Statement

All data were downloaded from the data repository from https://www.icpsr.umich.edu/web/DSDR/studies/37166. The current analyses used the publicly available dataset. The original protocol for the study is available:

Supplemental Material

Supplemental material for this article is available online.

References

SAMHSA. National substance use and mental health services survey, 2021; 2022. https://www.samhsa.gov/data/data-we-collect/n-sumhss-national-substance-use-and-mental-health-services-survey.

SAMHSA. Results from the 2013 national survey on drug use and health: Summary of national findings: DHHS; 2014.

Institute of Medicine Committee on Lesbian GB, Transgender Health I, Research G, Opportunities. The national academies collection: Reports funded by National Institutes of Health. The health of lesbian, gay, bisexual, and transgender people: Building a foundation for better understanding. Washington (DC): National Academies Press (US). Copyright © 2011, National Academy of Sciences; 2011.

Zhang

Qiao

Wharton

Liu

. Understand neighborhood-level stressor and resilience sources, healthcare discrimination and alcohol use among sexual minority groups in the United States: Insights from the all of us. Paper presented at: The Society for the Scientific Study of Sexuality Conference, San Diego, CA; 2024.

Hillyard

. LGBTQ+ drug and alcohol use: Discrimination breeds disparity. Br J Gen Pract. 2024;74(745):344–345. doi:10.3399/bjgp24X738813

Volpicelli

Menzies

. Rethinking unhealthy alcohol use in the United States: A structured review. Subst Abuse Res Treat. 2022;16:11782218221111832.

Meyer

. Minority stress and mental health in gay men. J Health Soc Behav. 1995;36(1):38–56. doi:10.2307/2137286

Chang

McMahon

Qiao

Zhang

. Discrimination, stress, and unhealthy alcohol and substance use among ethnoracial minority gay, bisexual, and other men who have sex with men. School of Nursing, University of Rochester; 2025.

Lewis

Romano

Ehlke

, et al. Minority stress and alcohol use in sexual minority women's daily lives. Exp Clin Psychopharmacol. 2021;29(5):501–510. doi:10.1037/pha0000484

10.

Evans

. Adding interactions to models of intersectional health inequalities: Comparing multilevel and conventional methods. Soc Sci Med. 2019;221:95–105. doi:10.1016/j.socscimed.2018.11.036

11.

Green

Evans

Subramanian

. Can intersectionality theory enrich population health research? Soc Sci Med. 2017;178:214–216. doi:10.1016/j.socscimed.2017.02.029

12.

Motley

Victorian

Denis

Brooks

. Applying an intersectionality framework to health services research. Fam Syst Health. 2023;41(4):417–424. doi:10.1037/fsh0000859

13.

Haardörfer

Windle

Berg

. Mental health and substance use among women and men at the intersections of identities and experiences of discrimination: Insights from the intersectionality framework. BMC Public Health. 2019;19(1):108. doi:10.1186/s12889-019-6430-0

14.

Evans

. Modeling the intersectionality of processes in the social production of health inequalities. Soc Sci Med. 2019;226:249–253. doi:10.1016/j.socscimed.2019.01.017

15.

DHHS. Healthy people 2030, social determinants of health; 2025. https://health.gov/healthypeople/objectives-and-data/social-determinants-health.

16.

Zhang

Dreisbach

Liu

. Analysis of social determinants of health among gay, bisexual and men who have sex with men in the United States: Insights from the all of us project. Paper presented at: LGBTQ Annual Conference, Altalan, GA; 2024.

17.

Yearby

. Structural racism and health disparities: Reconfiguring the social determinants of health framework to include the root cause. J Law Med Ethics. 2020;48(3):518–526. doi:10.1177/1073110520958876

18.

Lui

Martinez

Greenfiled

Kerr

. Quantification of racial and ethnic disparities in alcohol-related problems in light of different methodological approaches. Addiction. 2025;120(5):873–883. doi:10.1111/add.16755

19.

Pedrose

Pinheiro

Filho

de Araujo

. Machine learning-based prediction models for syphilis acquired during HIV pre-exposure prophylaxis. SSRN; 2024.

20.

Chow

EPF

, et al. A machine-learning-based risk-prediction tool for HIV and sexually transmitted infections acquisition over the next 12 months. J Clin Med. 2022;11(7):1818. doi:10.3390/jcm11071818

21.

Yang

Chen

, et al. XGBoost-SHAP-based interpretable diagnostic framework for Alzheimer's disease. BMC Med Inform Decis Mak. 2023;23(1):137. doi:10.1186/s12911-023-02238-9

22.

Balzer

Havlir

Kamya

, et al. Machine learning to identify persons at high-risk of human immunodeficiency virus acquisition in rural Kenya and Uganda. Clin Infect Dis. 2020;71(9):2326–2333. doi:10.1093/cid/ciz1096

23.

Goodman

Kaminsky

Lessler

. What is machine learning? A primer for the epidemiologist. Am J Epidemiol. 2019;188(12):2222–2239.

24.

Deo

. Machine learning in medicine. Circulation. 2015;132(20):1920–1930. doi:10.1161/CIRCULATIONAHA.115.001593

25.

Meyer

. Generations: A study of the life and health of LGB people in a changing society, United States, 2016-2019; 2023.

26.

Higgins-Biddle

Babor

. A review of the alcohol use disorders identification test (AUDIT), AUDIT-C, and USAUDIT for screening in the United States: Past issues and future directions. Am J Drug Alcohol Abuse. 2018;44(6):578–586. doi:10.1080/00952990.2018.1456545

27.

Bush

Kivlahan

McDonell

Fihn

Bradley

. The AUDIT alcohol consumption questions (AUDIT-C): An effective brief screening test for problem drinking. Ambulatory care quality improvement project (ACQUIP). Alcohol use disorders identification test. Arch Intern Med. 1998;158(16):1789–1795. doi:10.1001/archinte.158.16.1789

28.

Neufeld

Ullrich

Merchant

KAS

, et al. Change in drug use disorders identification test - Consumption (DUDIT-C) with telehealth treatment compared to in-person treatment. Subst Use Misuse. 2023;58(9):1168–1171. doi:10.1080/10826084.2023.2212038

29.

Hildebrand

. The psychometric properties of the drug use disorders identification test (DUDIT): A review of recent research. J Subst Abuse Treat. 2015;53:52–59. doi:10.1016/j.jsat.2015.01.008

30.

Roberts

Zhao

Verplaetse

, et al. Using machine learning to predict heavy drinking during outpatient alcohol treatment. Alcohol Clin Exp Res. 2022;46(4):657–666. doi:10.1111/acer.14802

31.

Afshar

. To err is machine: Considerations on the clinical impact of machine learning models in patients with unhealthy alcohol use. Alcohol Clin Exp Res. 2022;46(6):912–914. doi:10.1111/acer.14842

32.

Beretta

Santaniello

. Nearest neighbor imputation algorithms: A critical evaluation. BMC Med Inform Decis Mak. 2016;16(Suppl 3):74. doi:10.1186/s12911-016-0318-z

33.

Dong

Peng

. Principled missing data methods for researchers. SpringerPlus. 2013;2(1):222. doi:10.1186/2193-1801-2-222

34.

Luo

. Performance-weighted-voting model: An ensemble machine learning method for cancer type classification using whole-exome sequencing mutation. Quant Biol. 2020;8(4):347–358. doi:10.1007/s40484-020-0226-1

35.

Armeli

Feinn

Litt

Tennen

. Daily stress, drinking motives and alcohol co-use with other drugs. Subst Use Misuse. 2024;59(7):1059–1066. doi:10.1080/10826084.2024.2320374

36.

Dunbar

Siconolfi

Rodriguez

, et al. Alcohol use and cannabis use trajectories and sexual/gender minority disparities in young adulthood. Psychol Addict Behav. 2022;36(5):477–490. doi:10.1037/adb0000806

37.

West

Peasley-Miklus

Graham

, et al. Impact of alcohol and drug use on smoking and cessation in socioeconomically disadvantaged young adults. Addict Behav. 2020;110:106486. doi:10.1016/j.addbeh.2020.106486

38.

Squeglia

Jacobus

Tapert

. The effect of alcohol use on human adolescent brain structures and systems. Handb Clin Neurol. 2014;125:501–510. doi:10.1016/B978-0-444-62619-6.00028-8

39.

Forsythe

Pick

Tremblay

Malaviya

Green

Sandman

. Humanistic and economic burden of conversion therapy among LGBTQ youths in the United States. JAMA Pediatr. 2022;176(5):493–501. doi:10.1001/jamapediatrics.2022.0042

40.

The Lancet. Ending LGBT conversion therapies. Lancet. 2015;385(9977):1478. doi:10.1016/S0140-6736(15)60737-1

41.

Chang

Zhang

LeBlanc

. Application of minority stress theory to unhealthy alcohol use among ethnic minority gay, bisexual men, and other men who have sex with men. Paper presented at: The 2023 Asian American Pacific Islander Nurses Association Conference, Los Angeles, CA; 2023.

42.

AoU. Social determinants of health – All of us research hub; 2024. https://www.researchallofus.org/data-tools/survey-explorer/social-determinants-survey/.

43.

Donaldson

Stupplebeen

Couch

, et al. Perceived discrimination and youth vaping: The role of intersectional identities. Drug Alcohol Depend. 2024;260:111313. doi:10.1016/j.drugalcdep.2024.111313

44.

, et al. Web-Based risk prediction tool for an individual's risk of HIV and sexually transmitted infections using machine learning algorithms: Development and external validation study. J Med Internet Res. 2022;24(8):e37850. doi:10.2196/37850

Supplementary Material

Please find the following supplemental material available below.

For Open Access articles published under a Creative Commons License, all supplemental material carries the same license as the article it is associated with.

For non-Open Access articles published, all supplemental material carries a non-exclusive license, and permission requests for re-use of supplemental material or any part of supplemental material shall be sent directly to the copyright owner as specified in the copyright notice associated with the article.

0.00 MB

3.72 MB

0.03 MB

0.10 MB