Predictive Models to Identify Factors Associated with a High One-Time Sexual Partnership Rate Among U.S. Men Who Have Sex with Men

Abstract

One-time sexual partnerships have the potential to generate densely connected sexual networks. The distribution of one-time partnerships among MSM is often right-skewed, wherein a small proportion of MSM have a high number of partners. Identifying a set of covariates to predict who these individuals are can aid in developing and deploying STI prevention interventions and contribute to our understanding of STI transmission dynamics. We fit several negative binomial models to estimate the count of one-time partnerships among US MSM using data from a web-based egocentric network survey. Penalized logistic models were fit to evaluate predictors of MSM with high partnership counts at different binary cut points. Several covariates were significant predictors of a high one-time partnership rate in every model, including the number of main (partner considered a boyfriend, significant other, or life partner) and casual (an ongoing relationship, but not a main partner) partners, engaging in exchange sex, mobile app or internet use to meet men, and non-injection drug use. In the negative binomial models, the rate of one-time partnership formation was lower for those having one main partner compared to none, and higher for those with > 1 main partner, or any casual partners. Logistic models for partnership rates with a cut point at 6 partners per year considered as a high count had the overall best predictive performance. Overall, risk factors for HIV tend to occur together. We identified a set of covariates that were consistent predictors of high partnership rates. These models and predictors could be used as indicators for HIV/STI prevention interventions, such as partner-based services, and to aid in the parameterization of transmission models.

Introduction

Men who have sex with men (MSM) share a disproportionate burden of HIV and sexually transmitted infections (STIs). In 2022, 71% of all new HIV diagnoses occurred among MSM in the United States (U.S.), and in 2023, MSM accounted for 33% of reported primary and secondary syphilis cases [1, 2]. Yet MSM represent less than 5% of the total U.S. population [3-5]. Factors that play a role in this disparity include the high risk of HIV and STI transmission associated with anal intercourse, frequency of sex acts, and a high rate of sexual partnership formation [6-8]. One-time partnerships, which occur without duration (i.e., at a single point in time), can form at a higher rate than ongoing partnerships. This can increase both an individual’s own HIV/STI risk and their partner’s risk [9]. Some individuals are more likely to form one-time partnerships than others. Characterizing these individuals could aid in our understanding of what drives one-time partnership formation and in targeting HIV prevention efforts to high-risk individuals.

The potential effect of one-time partnerships on disease spread is two-fold. First, one-time partnerships can rapidly increase a person’s total number of partnerships over a short period of time, increasing their acquisition risk. Secondly, concurrency, or an overlap of partnerships in time, can increase a person’s partner’s HIV acquisition risk [10-12]. For this to occur, one-time partnerships would need to occur during an ongoing partnership. Most research has focused on the overall effect of concurrency, rather than on the pattern of an individual’s partnership types (ongoing vs. one-time) and how this affects transmission dynamics. Therefore, characterizing the association between ongoing partnerships and one-time partnerships would be useful for understanding if these partnerships occur together.

Observed partnership distributions (also called degree distributions) are typically right-skewed, where most individuals have a few partners at most, but a small subgroup has a much larger number of partners [13]. One-time partnerships may contribute substantially to the right tail of a degree distribution because of the potential to form at a higher rate than ongoing partnerships [14]. Degree distributions in which even a small proportion of individuals have many partnerships can result in a dense network with many sexual connections between individuals within the network [15]. The recent Monkeypox virus (MPXV) outbreak is a good example of how a small number of one-time sexual partnerships can disproportionately affect disease spread. In a model developed by the Centers for Disease Control and Prevention, half of the MPXV transmission among MSM was a result of one-time sexual partnerships, even though these partnerships accounted for < 3% of all partnerships on a given day [16]. Core group theory posits that the subset of the population with a high partnership formation rate is critical for maintaining the ongoing transmission of infections, as was shown with MPXV outbreak [16-18]. Core group theory has primarily been examined in heterosexual contexts, where a core group (e.g., sex workers) may propagate the epidemic into the general population [17]. Both the core group and a bridge population connected to the rest of the population are important actors for receiving interventions to limit the spread of disease [19]. A similar approach to deploying interventions to MSM could be used by targeting core group members (those with many one-time partnerships and concurrent ongoing partners) and their partners for interventions such as what is done with expedited partner therapy.

The sociological mechanisms that generate the observed right-skewed degree distribution are not well described for MSM sexual partnerships. Early research focused on the preferential attachment model, where individuals with many partners attract more partners [20-22]. Preferential attachment can be modeled using a power-law distribution, which can account for a right-skewed distribution. However, epidemic simulations using this distribution were unable to accurately predict epidemic outcomes [23]. More recent research identified temporal variation in individual one-time partnership rates among MSM, suggesting that people may have seasons of risk over time [24]. Understanding what factors predict the degree or seasons of high degree can aid in developing more accurate mathematical and mechanistic models.

Differences in degree by behavioral and demographic characteristics have been observed [25, 26]. For example, higher casual partnership counts were reported by White MSM, those reporting exchange sex, and those reporting any drug use. Drug use has consistently been identified as an important predictor of high-risk sexual behavior, including having multiple sex partners [25, 27]. Adverse mental health and stigma have also been linked to having multiple sex partners and also substance abuse [28]. Several clinical prediction models have been developed to evaluate a person’s risk of HIV acquisition, where a recent high number of partners was an important predictor [7, 8]. Determining what factors are associated with specifically one-time partnership rates can aid in targeting interventions to high-risk individuals and their partners. One-time partnership rates are relevant for understanding disease transmission because they contribute to the epidemic potential across a network and indicate individual HIV/STI transmission risk [9-11, 15]. Therefore, a set of covariates that could predict one-time partnership rates could be used to improve our understanding of network-level risks.

In this study, we sought to (1) model the distribution of one-time partnership rates and (2) identify a set of covariates that can predict who is likely to have a high number of partners. Our goal was not to evaluate causality between the predictors and one-time partnership rates, but rather to assess if they can predict the occurrence of such partnerships. Our results can aid in designing, optimizing, and delivering prevention tools for STIs, such as partner services, an intervention that critically depends on the number of current or recent partnerships [29].

Methods

Study Design

Data used in this analysis were from ARTnet, a web-based egocentric network study of U.S. MSM conducted between 2017 and 2019 [30]. Recruitment occurred via the American Men’s Internet Survey (AMIS), a national web-based survey about the sexual health of MSM. Respondents were asked to participate in ARTnet once they completed AMIS [31]. Eligibility for ARTnet included male sex at birth, a current male gender identity, lifetime history of sexual activity with another man, and age between 15 and 65. For ARTnet, data were collected on the sexual partnerships of each respondent; the partners of the respondents were not surveyed directly. The final sample size of ARTnet, after deduplication within and across survey waves, was 4,904 participants, who reported on 16,198 sexual partnerships. The Emory University Institutional Review Board approved the study.

Measures

Male sexual partnerships were defined by type: main (partner considered a boyfriend, significant other, or life partner), casual (an ongoing relationship, but not a main partner), and one-time contacts (no relational persistence). In the 2017 survey, participants reported the number of main, casual, and total partners they had in the 12 months prior to the study, along with detailed information on their five most recent partnerships. In the 2018 survey, participants additionally reported on the number of one-time partnerships in the past 12 months. Because the count of one-time partners was asked directly only in 2018, we estimated this variable’s value by subtracting the counts of main and casual partnerships from the total count of partnerships. For consistency, we did this for both years and checked how well the 2018 estimates matched the measured variable. The mean difference between the derived and measured variable for 2018 was 0.7, and the median was 0.0. As a secondary outcome, we created cut points representing high counts of one-time partnerships. We evaluated cut points at 6 (on average, one new partner every two months), 12 (on average, one new partner per month), and 24 (on average, two new partners per month) partners to derive this variable.

Four categories of main effect predictors were considered: demographic (age, race, education level, census division, and urbanicity), clinical (reported HIV status and reported PrEP use in the past 12 months), network degree (main, casual, or total partnership degree), and behavioral. The degree term (main, casual, or total) included in the model varied by the outcome: for cut points at 12 and 24 partners per year, total degree was included as a predictor, which was the sum of main and casual degree. All other outcomes used main and casual degree as independent predictors. Behavioral variables included: whether the participant had revealed to anyone that he is attracted to men, sexual identity, ever injected drugs, ever experienced stigma from friends and family, healthcare-related stigma, general social stigma, and a report of the following in the past 12 months: sex with a female partner, exchange sex (exchanging drugs or money for sex), use of apps or internet sites to meet men, non-injection drug use, experiencing any stigma, and homelessness.

We also evaluated six interaction terms: age × HIV status, age × app use, race × HIV status, sexual identity × sex with females, age × main degree, and age × casual degree. Interaction terms were included to assess whether they improved model fit and were chosen based on previous research. The former four were previously found to be predictive of casual male partnership counts [25, 32]. For the age-by-degree terms, previous work suggested that partnership counts decrease with age [33]. We hypothesized that as an individual ages, they might have more ongoing partnerships, which in turn might result in fewer one-time partnerships.

Statistical Analyses

The distribution of each variable was examined prior to regression modeling, including summary statistics for the primary outcome within predictor categories. Stochastic regression imputation was used to impute values for all missing predictors using the MICE package in R [34]. Each imputation model included all predictors and the outcome. Logistic regression models were used for binary predictors and Bayesian polytomous regression for categorical variables with more than two levels. Using the imputed dataset, we fit negative binomial models for the primary outcome to account for the right-skewed nature of the degree distribution and logistic models for the secondary outcomes. Analyses were conducted in R 4.0.4. Analysis code may be found at: [https://github.com/EpiModel/PredPartRates].

We evaluated goodness of fit measures (described below) and measures of association (incident rate ratios or odds ratios) for several prespecified models, summarized in Table 1: (1) a full model including all predictors and six interaction terms (age × HIV status, age × app use, race × HIV status, sexual identity × sex with females, age × main degree, and age × casual degree), (2) a behavioral model with only the demographic, behavioral, and network predictors and the four relevant interaction terms, (3) a model with all predictors but no interaction terms, (4) a behavioral model with only the demographic, behavioral, and network predictors and no interaction terms, and (5) a reduced model for the binary cut point outcomes. We estimated the models without the clinical variables because we may wish to predict one-time partner rates prior to receipt of care. All models and predictors were prespecified based on a review of the relevant literature.

Table 1.

Description of prespecified models examined using negative binomial and logistic regression models

	Negative binomial models	Logistic models1
Model 1: All interactions - full	Demographic main effects2 Behavioral main effects3 Clinical main effects 4 Network main effects5 Age-squared term 6 Interactions: Age and HIV status Age and app use Race and HIV status Sexual identity and sex with females Age and main degree Age and casual degree	Demographic main effects Behavioral main effects Clinical main effects Network main effects Age-squared term 6 Interactions: Age and HIV status Age and app use Race and HIV status Sexual identity and sex with females Age and main degree Age and casual degree
Model 2: All interactions - behavioral	Demographic main effects Behavioral main effects Network main effects Age-squared term 4 Interactions: Age and app use Age and main degree Age and casual degree Sexual identity and sex with females	Demographic main effects e42b5 Behavioral main effects Network main effects Age-squared term 4 Interactions: Age and app use Age and main degree Age and casual degree Sexual identity and sex with females
Model 3: No interactions -full	Demographic main effects Behavioral main effects Clinical main effects Network main effects Age-squared term	Demographic main effects Behavioral main effects Clinical main effects Network main effects Age-squared term
Model 4: No interactions - behavioral	Demographic main effects Behavioral main effects Network main effects Age-squared term	Demographic main effects Behavioral main effects Network main effects Age-squared term
Model 5: Reduced6	Model not considered	Main effects terms include: Age Race Urbanicity Sexual identity Exchange sex Female sex partner App use Ever injected drugs Non-injection drug use Total degree

¹Logistic models fit using both regular and penalized regression

²Demographic effect predictors include: age, race/ethnicity, education level, census division, urbanicity

³Behavioral main effect predictors include: sexual identity, exchange sex with a male partner in the past 12 months, sex with females in past 12 months, used apps to meet men in past 12 months, out as attracted to men to anyone, ever injected drugs, non-injection drug use in the past 12 months, homelessness in past 12 months, experienced any sexual identity stigma in past 12 months, ever experienced any sexual identity stigma from friends and family, ever experienced health-related sexual identity stigma, ever experienced general social stigma related to sexual identity, casual male partner degree, main male partner degree

⁴Clinical main effect predictors include reported HIV status, PrEP use in the past 12 months

⁵Network main effect predictors include casual male partner degree, main male partner degree

⁶Reduced models specified to account for the smaller effective sample size in the higher cut point logistic models (cut points 12 and 24) to decrease the degrees of freedom

We estimated penalized logistic models for the binary outcomes to account for potential overfitting. Ten-fold cross-validation was used to derive the optimum penalty parameter. AIC and likelihood ratios were used to compare nested models. For the negative binomial models, we evaluated root mean squared error (RMSE) and mean absolute error (MAE), two measures of model fit that estimate the model’s average prediction error. Nagelkerke’s R² and the c-statistic (area under the curve) were estimated for the logistic models. Calibration plots (observed versus predicted outcomes) were generated for all models. Wald statistics were examined for each model to assess statistically significant predictors across models at an alpha level of 0.10.

To assess the magnitude of the association between the predictors and one-time partnership rates, we estimated incident rate ratios (IRRs) from two models: Model 3 (full model, no interactions) and Model 4 (behavioral model, no interactions). We chose these models for several reasons, including ease of interpretation and because Model 3 was the best-fitting model, and Model 4 was the best-fitting model among the behavioral models (Models 2 and 4). We used incident rate ratios because one-time partnerships occur at a single time point and are therefore incident events. We also estimated odds ratios from the best-fitting logistic model (Supplemental Table 6). Our main goal was not to interpret specific incidence rate ratios, but to assess the overall predictive capability of the covariates regarding the rate of one-time partnerships.

Bootstrap resampling was used to generate optimism-corrected estimates of the RMSE, MAE, R², and c statistic. Optimism occurs when the prediction performance is overestimated in the sample used to derive the model [32, 35]. For this analysis, 200 bootstrapped samples were generated, and each model was fit to each sample. The newly fitted models were then evaluated using the original sample. Optimism was calculated for each sample as the difference between performance in the bootstrap sample and in the original sample. The average optimism in the 200 samples was then subtracted from the model results in the original analysis to estimate the optimism-corrected performance measure.

Results

Descriptive statistics of the study participants before imputation are provided in Table 2, including the percent missing for each predictor. The study population was predominantly White (72%), college-educated (56%), and living in a large central metro (43%). The reported HIV prevalence in the sample was 9%, and 16% reported using PrEP in the past 12 months. Overall, the mean number of one-time partnerships in the sample was 4.6 partners/year, but ranged from 0 to 250, with 46% reporting no one-time partnerships in the past 12 months. The median number of one-time partners was 1.0 (IQR: 0.0, 3.0). Figure 1 shows the distribution of one-time partnerships in the sample. The mean count of one-time partnerships was highest among HIV-positive participants (8.6 partners/year), PrEP users (11.6 partners/year), those reporting any exchange sex in the past year (12.3 partners/year), ever injecting drugs (10.1 partners/year), two or more main partnerships in the past 12 months (8.0 partners/year to 21.5 partners/year), and two or more casual partnerships in the past 12 months. (9.9 partners/year to 24.6 partners/year).

Table 2.

Descriptive statistics of ARTnet study participants and negative binomial model results for the count of one-time partnerships in the past 12 months

	n (%)	Mean one-time partner count (SD)	Median one-time partner count (IQR)	IRR1 (95% CI)	IRR2 (95% CI)
Total	4904	4.6 (12.7)	1.0 (0.0, 3.0)	–	–
Demographic variables
Age
15–24	1324 (27.0)	3.0 (8.1)	1.0 (0.0, 3.0)	Ref	Ref
25–34	1268 (25.9)	4.2 (10.9)	1.0 (0.0, 3.0)	0.85 (0.73, 0.98)	0.97 (0.83, 1.13)
25–44	694 (14.2)	5.8 (16.4)	1.0 (0.0, 5.0)	0.94 (0.80, 1.12)	1.19 (0.99, 1.42)
45–54	833 (17.0)	6.2 (16.5)	1.0 (0.0, 4.0)	0.94 (0.80, 1.12)	1.12 (0.94, 1.33)
55–64	785 (16.0)	4.9 (13.0)	0.0 (0.0, 3.0)	0.78 (0.66, 0.94)	0.91 (0.76, 1.09)
Race/ethnicity
Black	266 (5.4)	3.5 (9.8)	1.0 (0.0, 3.0)	Ref	Ref
Hispanic	676 (13.8)	4.4 (10.8)	1.0 (0.0, 4.0)	1.07 (0.82, 1.39)	1.07 (0.81, 1.40)
Other	439 (9.0)	3.7 (8.0)	1.0 (0.0, 3.0)	0.97 (0.73, 1.28)	0.91 (0.68, 1.22)
White	3523 (71.8)	4.8 (13.7)	1.0 (0.0, 3.0)	1.00 (0.79, 1.26)	1.02 (0.79, 1.30)
Education level
Less than high school	170 (3.5)	2.4 (8.8)	0.0 (0.0, 1.0)	Ref	Ref
High school diploma or equivalent	446 (9.1)	3.1 (8.8)	1.0 (0.0, 2.0)	1.21 (0.86, 1.68)	1.19 (0.83, 1.69)
Some college or equivalent	1519 (31.0)	3.9 (11.7)	1.0 (0.0, 3.0)	1.32 (0.97, 1.79)	1.38 (0.99, 1.90)
College degree or higher	2742 (55.9)	5.3 (13.9)	1.0 (0.0, 4.0)	1.40 (1.02, 1.91)	1.68 (1.20, 2.33)
Missing3	27 (0.6)	5.3 (6.3)	1.0 (0.0, 3.0)	–	–
Census division
New England	250 (5.1)	4.4 (11.4)	1.0 (0.0, 4.0)	Ref	Ref
Middle Atlantic	632 (12.9)	4.7 (12.1)	1.0 (0.0, 3.0)	0.86 (0.66, 1.11)	0.82 (0.62, 1.08)
East North Central	698 (14.2)	4.4 (12.0)	1.0 (0.0, 3.0)	0.89 (0.69, 1.16)	0.85 (0.64, 1.11)
West North Central	296 (6.0)	4.3 (10.1)	1.0 (0.0, 3.0)	0.77 (0.57, 1.04)	0.75 (0.55, 1.03)
South Atlantic	1062 (21.7)	4.4 (13.5)	1.0 (0.0, 3.0)	0.92 (0.72, 1.17)	0.93 (0.72, 1.20)
East South Central	222 (4.5)	3.2 (9.5)	0.0 (0.0, 2.0)	0.75 (0.54, 1.04)	0.67 (0.48, 0.94)
West South Central	498 (10.2)	4.5 (14.1)	1.0 (0.0, 4.0)	0.85 (0.65, 1.12)	0.81 (0.61, 1.08)
Mountain	404 (8.2)	3.8 (9.5)	1.0 (0.0, 3.0)	0.87 (0.65, 1.15)	0.80 (0.59, 1.07)
Pacific	842 (17.2)	5.6 (14.9)	1.0 (0.0, 4.0)	0.78 (0.60, 1.01)	0.85 (0.65, 1.10)
Urbanicity
Large Central Metro	2116 (43.1)	5.6 (14.2)	1.0 (0.0, 5.0)	Ref	Ref
Large Fringe Metro	1056 (21.5)	4.3 (13.2)	1.0 (0.0, 3.0)	0.82 (0.72, 0.95)	0.73 (0.64, 0.85)
Medium Metro	921 (18.8)	3.5 (10.0)	0.0 (0.0, 2.0)	0.79 (0.69, 0.92)	0.71 (0.61, 0.82)
Small Metro	421 (8.6)	3.4 (10.0)	1.0 (0.0, 3.0)	0.84 (0.69, 1.03)	0.72 (0.59, 0.89)
Micropolitan	230 (4.7)	3.0 (9.0)	0.0 (0.0, 2.0)	0.86 (0.67, 1.11)	0.65 (0.50, 0.85)
Nonecore	160 (3.3)	3.9 (12.2)	0.0 (0.0, 1.5)	0.90 (0.67,1.22)	0.75 (0.56, 1.03)
Clinical variables
Reported HIV status
Negative	3726 (76.0)	4.7 (12.6)	1.0 (0.0, 4.0)	Ref	–
Positive	428 (8.7)	8.6 (19.2)	1.0 (0.0, 8.0)	2.00 (1.67, 2.43)	–
Unknown	750 (15.3)	1.6 (5.9)	0.0 (0.0, 1.0)	–	–
PrEP used in past 12 months
No	3318 (67.7)	2.4 (6.8)	0.0 (0.0, 2.0)	Ref	–
Yes	788 (16.1)	11.6 (21.9)	4.0 (1.0, 13.0)	3.38 (2.96, 3.87)	–
Missing	428 (8.7)	8.6 (19.2)	1.0 (0.0, 8.0)	–	–
Behavioral variables
Sexual identity
Heterosexual or straight	39 (0.8)	3.9 (15.4)	0.0 (0.0, 2.0)	Ref	Ref
Homosexual or gay	3982 (81.2)	4.6 (12.9)	1.0 (0.0, 3.0)	0.87 (0.45, 1.60)	0.87 (0.43, 1.68)
Bisexual	775 (15.8)	4.4 (12.2)	1.0 (0.0, 3.0)	0.90 (0.47, 1.64)	0.81 (0.40, 1.54)
Other sexual identity	59 (1.2)	2.9 (6.7)	0.0 (0.0, 3.0)	0.80 (0.36, 1.73)	0.80 (0.34, 1.81)
Missing	49 (1.0)	3.9 (11.2)	0.0 (0.0, 1.0)	–	–
Exchanged sex with a male partner in past 12 months
No	4516 (92.1)	4.0 (11.5)	1.0 (0.0, 3.0)	Ref	Ref
Yes	314 (6.4)	12.3 (22.4)	3 0.0 (1.0, 14.0)	1.95 (1.60, 2.38)	1.83 (1.49, 2.26)
Missing	74 (1.5)	4.5 (12.3)	1.0 (0.0, 3.0)	–	–
Sex with females in past 12 months
No	4301 (87.7)	4.3 (12.4)	1.0 (0.0, 3.0)	Ref	Ref
Yes	546 (11.1)	6.2 (15.4)	1.0 (0.0, 4.0)	1.38 (1.16, 1.66)	1.37 (1.14, 1.66)
Missing	57 (1.2)	4.4 (8.1)	1.0 (0.0, 3.0)	–	–
Used apps to meet men in past 12 months
No	614 (12.5)	0.4 (2.6)	0.0 (0.0, 0.0)	Ref	Ref
Yes	3252 (66.3)	5.6 (14.4)	1.0 (0.0, 5.0)	7.13 (5.94, 8.58)	9.15 (7.59, 11.03)
Missing	1038 (21.2)	3.7 (9.8)	0.0 (0.0, 2.0)	–	–
Out to anyone as attracted to men
No	177 (3.6)	3.5 (9.5)	1.0 (0.0, 2.0)	Ref	Ref
Yes	4535 (92.5)	4.6 (12.9)	1.0 (0.0, 3.0)	1.09 (0.81, 1.44)	1.28 (0.93, 1.73)
Missing	192 (3.9)	4.2 (9.5)	1.0 (0.0, 3.0)	–	–
Ever injected drugs
No	4673 (95.3)	4.3 (12.0)	1.0 (0.0, 3.0)	Ref	Ref
Yes	216 (4.4)	10.1 (23.1)	1.0 (0.0, 8.3)	1.24 (0.97, 1.61)	1.42 (1.10, 1.85)
Missing	15 (0.3)	2.1 (3.9)	1.0 (0.0, 2.5)	–	–
Non–injection drug use in past 12 months
No	3083 (62.9)	3.6 (11.5)	0.0 (0.0, 2.0)	Ref	Ref
Yes	1789 (36.5)	6.2 (14.5)	1.0 (0.0, 5.0)	1.34 (1.20, 1.49)	1.36 (1.21, 1.52)
Missing	32 (0.7)	3.4 (5.2)	1.0 (0.0, 4.0)	–	–
Homelessness in Past 12 Mo.
No	4263 (86.9)	4.4 (12.4)	1.0 (0.0, 3.0)	Ref	Ref
Yes	324 (6.6)	6.5 (17.6)	1.0 (0.0, 5.0)	1.14 (0.94,1.40)	1.17 (0.96, 1.45)
Missing	317 (6.5)	4.6 (11.0)	1.0 (0.0, 4.0)	–	–
Experienced any stigma in past 12 months
No	3654 (74.5)	4.3 (12.1)	1.0 (0.0, 3.0)	Ref	Ref
Yes	1099 (22.4)	5.5 (15.0)	1.0 (0.0, 4.0)	1.22 (1.08, 1.38)	1.22 (1.08, 1.39)
Missing	151 (3.1)	3.1 (7.4)	1.0 (0.0, 2.0)	–	–
Ever experienced stigma from friends & family
No	2069 (42.2)	4.2 (11.0)	1.0 (0.0, 3.0)	Ref	Ref
Yes	2754 (56.2)	4.8 (13.9)	1.0 (0.0, 3.0)	0.98 (0.88, 1.09)	0.98 (0.88, 1.10)
Missing	81 (1.7)	5.1 (11.8)	1.0 (0.0, 4.0)	–	–
Ever experienced healthcare–related stigma
No	3493 (71.2)	4.3 (12.6)	1.0 (0.0, 3.0)	Ref	Ref
Yes	1376 (28.1)	5.1 (13.1)	1.0 (0.0, 4.0)	1.18 (1.05, 1.32)	1.09 (0.97, 1.23)
Missing	35 (0.7)	2.2 (3.3)	1.0 (0.0, 2.5)	–	–
Ever experienced general social stigma
No	1751 (35.7)	3.9 (10.7)	1.0 (0.0, 3.0)	Ref	Ref
Yes	3149 (64.2)	4.9 (13.7)	1.0 (0.0, 3.0)	0.86 (0.77, 0.97)	0.92 (0.81, 1.03)
Missing	4 (0.1)	1.0 (1.4)	0.5 (0.0, 1.5)	–	–
Main partnership degree
0	2988 (60.9)	5.0 (12.8)	1.0 (0.0, 4.0)	Ref	Ref
1	1835 (37.4)	3.5 (12.1)	0.0 (0.0, 2.0)	0.76 (0.68, 0.85)	0.84 (0.75, 0.94)
2 to 6	81 (1.7)	8.8 (17.8)	2.0, (0.0, 9.0)	1.27 (0.89, 1.87)	1.38 (0.95, 2.07)
Casual partnership Degree
0	3229 (65.8)	2.2 (8.0)	0.0 (0.0, 2.0)	Ref	Ref
1	946 (19.3)	6.0 (12.4)	2.0 (0.0, 6.0)	1.87 (1.65, 2.13)	2.14 (1.88, 2.45)
2	400 (8.2)	9.9 (18.0)	3.0 (0.0, 11.0)	2.56 (2.16, 3.06)	3.09 (2.59,3.72)
3	205 (4.2)	14.3 (20.8)	7.0 (2.0, 16.0)	3.62 (2.88, 4.62)	4.57 (3.60, 5.89)
4	87 (1.8)	18.4 (31.8)	8.0 (1.0, 20.0)	5.76 (4.10, 8.35)	7.76 (542, 11.50)
5	37 (0.8)	24.6 (38.0)	15.0 (4.0, 25.0)	7.8 (4.72, 13.98)	11.67 (6.93 (21.40)

¹IRRs calculated from Model 3 with age categorized

²IRRs calculated from Model 4 and age categorized

³Proportion missing shown for each predictor; for regression output (incident rate ratios), missing values imputed using stochastic regression imputation

⁴Highlighted variables were significant for every model and outcome

Fig. 1.

The distribution of observed one-time partnership counts of US MSM in the ARTnet study 2017–2018 (n = 4,904)

Several predictors were statistically significant across outcomes (count of one-time partnerships in the past year or high vs. low one-time partner count) and models (Supplemental Tables 2-5 and highlighted in Table 2). Having a higher degree was generally associated with more one-time partners. Furthermore, having one main partner was associated with fewer one-time partners than having no main partners. A report of exchange sex, app use, and non-injection drug use was statistically significantly associated with all outcomes. Other routinely significant predictors across outcomes and models included education (higher education levels associated with higher one-time partner count), living outside of a large metro area, having had female sex partners, having ever injected drugs, having experienced stigma in the past 12 months, HIV-positivity, and PrEP use. Among the interaction terms, an age by degree term was significant in every model. None of the other interaction terms were consistently significant across outcomes and models.

Incidence rate ratios (IRRs) for the rate of one-time partnership formation and their confidence intervals (CIs) are shown in Table 2. For ease of interpretation, age was converted into a categorical variable here. Individuals reporting app use to meet men had an estimated rate ratio of 9.2 (95% CI: 7.6, 11.0). The rate ratios among those reporting exchange sex (IRR = 1.8, 95% CI: 1.5, 2.3), sex with a female partner (IRR = 1.4, 95% CI: 1.1, 1.7), ever injecting drugs (IRR = 1.4, 95% CI: 1.1, 1.9), and non-injection drug use (IRR = 1.4, 95% CI: 1.2, 1.5) were also significantly above one compared to those who did not report these behaviors. Among the network variables, having a single main partnership resulted in a lower one-time partnership rate than having no main partnership (IRR = 0.84, 95% CI: 0.75, 0.94). Conversely, having any casual partnerships resulted in a higher rate of one-time partnerships (IRRs ranged from 2.1 to 11.7, going from 1 to 5 casual partnerships in the past 12 months, with each successive number of casual partnerships resulting in a higher rate ratio).

Supplemental Table 1 shows summary results for each model (model details specified in Table 1). Models with lower RMSE and MAE and higher AIC suggest a better fit to the data. Among the negative binomial models, Model 4 (the behavioral model with no interaction terms) had the highest AIC. However, Model 3 (the clinical model with no interaction terms) had the lowest RSME and MAE. Table 3 shows the mean observed count of one-time partnerships, compared with the mean predicted count, for each model, stratified by observed one-time partnership count (outcome group). For observed counts between 0 and 7, the models tended to overpredict, and for counts over 8, they tended to underpredict. This can also be seen in Fig. 2, which shows the observed one-time partner count distribution (shown in gray) plotted against the predicted distribution for each model. The top row is zoomed in to counts less than 25, and the bottom row is zoomed in to counts of 25 and up. Every model underpredicts the count of zeros and overpredicts counts less than approximately 10. Calibration plots for the negative binomial models can be found in Supplemental Figs. 1-9.

Table 3.

Predicted versus observed mean counts of one-time male partnerships in the past 12 months

	n (%)	Observed	Model 1	Model 2	Model 3	Model 4
Overall	4904 (100)	4.54	5.07	4.84	4.94	4.80
Outcome group1
0	2256 (46)	0.00	2.47	2.90	2.54	2.95
1	802 (16)	1.00	3.34	3.67	3.35	3.67
2	429 (9)	2.00	4.83	4.85	4.86	4.86
3	249 (5)	3.00	4.84	4.76	4.86	4.89
4 to 5	283 (6)	4.46	7.53	6.67	7.12	6.60
6 to 7	173 (4)	6.49	8.28	6.85	7.97	6.73
8 to 12	241 (5)	9.59	9.17	7.89	9.52	8.08
13 to 21	228 (5)	16.42	11.94	9.86	12.02	10.09
22+	243 (5)	47.08	19.92	15.51	16.87	13.83

¹Groups based on 5% quantiles

Fig. 2.

The distribution of predicted vs. observed one-time partnership counts for each negative binomial model. The top row shows the distribution for partnership counts ≤ 24 and the bottom row shows the distribution for partnership counts > 24. The grey-filled bars are the observed counts, and the color-filled bars are the predicted counts

For the logistic models, model performance varied by the cut point used to derive the dichotomous outcome. Among the models with the clinical variables, Model 3 performed best at every cut point, with the overall best performance at a cut point of 6 partners (optimism-corrected R2 = 0.32 and c-statistic = 0.82). Model 3, with a cut-point of 24, had the highest c-statistic (0.84) but a mediocre R² (0.24). Odds ratios and 95% confidence intervals for Model 3, cut-point 6, can be found in the supplemental appendix (Supplemental Table 6).

Discussion

In this study, we modeled one-time partnership rates among U.S. MSM to identify factors that predict a high number of one-time partners. Because one-time partnership formation can occur at a high rate, identifying individuals likely to have many one-time partners could be beneficial for HIV prevention efforts. Furthermore, behavior changes rapidly at both the individual and population levels, so understanding when someone might be in a high-risk period is important to target interventions appropriately [24, 36, 37].

Several behavioral, network, and clinical predictors were consistently significant across models and outcomes. These included a report of exchange sex, use of apps to meet men, a report of non-injection drug use, higher network degree, and PrEP use. Logistic models where partnership rates were modeled as a binary outcome of high versus low rates fit best with c statistics ranging from 0.77 to 0.84, whereas none of the negative binomial models fit the data well. Future research could evaluate new ways of modeling the distribution of one-time partnerships using the predictors we identified as important.

We identified several significant predictors of one-time partnership rates among U.S. MSM. The behavioral predictors included exchange sex, app use to meet men, and non-injection drug use, all of which have consistently been associated with multiple sex partners in the literature. Here, we identified their association with one-time partnership rates, which, to our knowledge, has not previously been demonstrated. In our study, the mean one-time partnership rate was 14 times higher among those reporting app use to meet men compared with those reporting no app use. Additionally, app use was common in our sample, with over two-thirds reporting app use over the past year. Previous research has linked app use, chat rooms, and other forms of meeting partners online to having multiple sex partners. For example, in a study among users of Grindr, an app to find partners for MSM, the mean number of both insertive and receptive anal sex partners in the past six months was approximately five partners [38]. One study showed that online partnerships may not be an independent HIV risk but are instead associated with other behaviors that increase risk, such as multiple sex partners [39]. Exchange sex (e.g., exchanging drugs or money for sex) was also an important predictor in our analysis. The rate of one-time partnerships was almost twice as high among those reporting exchange sex compared to those not reporting exchange sex. Similar to our results for one-time partnerships, previous research found a two-fold increase in the median number of casual partnerships among those reporting exchange sex versus no exchange sex, suggesting that those engaging in exchange sex may be more likely to have multiple partners of varying types [25]. Finally, non-injection drug use was significant in every model, and injection drug use was significant in most models. Research has implicated both behaviors in having multiple sex partners. In one study, the median number of casual partnerships among those reporting non-injection drug use was almost twice as high as those reporting no drug use (4.3 vs. 2.7 casual partnerships) [25]. Injection drug use may not have been as strongly associated with our outcomes as non-injection drug use because we asked about ever injecting drugs rather than recent injection. These results indicate that risky behaviors travel together and that a report of one or more may indicate that someone has a high one-time partnership rate. Furthermore, generally, the more partnerships someone has, the higher the HIV risk, although this depends on many additional factors as well, such as condom use, PrEP use, etc.

The network degree variables were also important predictors in all our models. Our results indicate that those reporting a high one-time partnership rate were also more likely to have multiple ongoing sex partners. Specifically, having multiple persistent main partners or casual partners was associated with a higher rate of one-time partnerships compared with having one main or no casual partners. For example, the one-time partnership rate among those reporting five casual partners was almost eight times higher than that of those reporting no casual partners. Having more than one casual partner may be common among MSM: results from one study reported that over three-quarters of participants had a casual partner, and almost half had both main and casual partners in the past year [25, 40]. Good predictive models are necessary to better understand what factors are associated with having multiple partnerships of different types.

Network studies show that partnership concurrency (2 or more ongoing partnerships) can facilitate HIV and STI transmission by increasing network connectivity [12, 14]. Concurrency results in an increased risk to the partners of highly connected individuals, particularly partners who engage in repeated sex acts with someone who has a high rate of partnership formation. One question is whether individuals engage in coital dilution, which occurs when someone reduces their frequency of sex acts when they form an additional partnership. In this situation, the effect of concurrency on the epidemic potential of a network may be unchanged or reduced [41, 42]. If there were coital dilution in a network, we might expect a negative relationship between additional main partnerships and casual or one-time partners. However, we observed the opposite, suggesting that there may be some underlying topology for having more of all partnership types. Future research should explore this idea further and whether such individuals are good candidates for targeting HIV and STI interventions.

A secondary goal in our analysis was to evaluate using the negative binomial distribution to model one-time partnership rates. The best-fitting models were those without the interaction terms, suggesting that the assessed interactions may not have a substantial impact on predicting one-time partnership rates. Overall, none of the models fit the data well; zeroes were overpredicted, and high rates were underpredicted. However, when we dichotomized the rate into high and low, the logistic models fit moderately well, with c-statistics ranging from 0.77 to 0.84. Previous work hypothesized that a preferential attachment model, underpinned by a power-law distribution, might be a good fit for modeling degree [23]. Under this model, the probability of partner selection is a positive function of the degree of a potential partner; individuals with higher degrees attract more partners. However, this seems like an unlikely mechanism of sexual partnership formation, and indeed, using a power-law distribution consistently resulted in inaccurate epidemic predictions. This result, along with our work here, indicates that a single regression model does not adequately capture the partner selection process. One process might result in someone having any or no one-time partners, and another might result in someone having many one-time partners. As noted above, there appear to be different partnership-formation typologies that should be examined in more detail in future work.

Limitations

Both the primary and secondary outcomes we evaluated were based on a derived variable, which could result in information bias. The derived variable was calculated as the difference between total partnerships and the count of ongoing partnerships. In our data, the number of reported persistent partnerships (main and casual partnerships) in the past 12 months was restricted to the five most recent in the previous year. It is possible that some individuals had more than five persistent partnerships and, therefore, fewer estimated one-time partnerships. However, only 37 respondents reported at least five persistent partnerships, so the resulting overestimate of one-time partnerships would likely be small. Our data are cross-sectional, so we could not establish temporality between the predictors and the outcomes. Although our goal was not to evaluate causality of the predictors on the outcome, to be useful as indicators of current or future high one-time partnership counts, the predictors would need to occur right before or simultaneous to a period of one-time partner acquisition. For this reason, we chose to fit models both with and without the clinical variables. Finally, our data included a small proportion of minority race/ethnicities who are historically underrepresented in online surveys [43]. For instance, our sample was comprised of 72% White MSM and only 5% Black MSM. Because of the small proportions in non-White categories, we may not have had the power to detect a strong association between race/ethnicity and one-time partnerships rates. Future work should consider alternative means of recruitment to ensure an adequate sample of minority race/ethnicities.

Conclusions

We identified several behavioral and network predictors of one-time partnership rates that can be used as indicators for HIV and STI interventions, such as partner-based services. For example, those reporting app use to meet partners, or any non-injection drug use could be targeted for testing and partner services. Although the negative binomial models poorly predicted the distribution of one-time partnership rates, the logistic models had good discriminative ability. Future research should evaluate new ways of modeling the distribution of one-time partnerships, such as modeling different regions of the distribution separately and incorporating additional behavioral variables. Additionally, qualitative methods could aid our understanding of how individuals choose their partners.

Supplementary Material

Supplementary Information The online version contains supplementary material available at https://doi.org/10.1007/s10461-026-05107-4.