Data were collected at baseline from January 2018 to December 2019 from the Chicago site of the multisite N2 cohort study, a prospective longitudinal study. Participants were recruited from a community health clinic and via peer referral in the South Side of Chicago, Illinois, United States. The community health clinic provides next-generation testing and notification, integrated biobehavioral HIV prevention, and community mobilization with the aim of eliminating new HIV transmission in the United States by 2041. The recruitment strategy was designed to leverage clients of the clinic and participants from existing research projects. We used a probabilistic variant of peer referral sampling that allowed participants to refer up to 6 members of their social networks. Participant inclusion criteria were the following: (1) Black or African American racial identity, (2) 16 years old or older, (3) assignment of male sex at birth, (4) residency in the Chicago Metropolitan Statistical Area (MSA), (5) having at least one sexual encounter with a cisgender man or transgender woman in the past year, (6) willingness to wear a GPS device, and (7) intention to remain in the Chicago MSA during the course of the study (~2 years). Additional information about the parent N2 study can be found elsewhere [32]. Given that the data in this study were collected 6 years ago, this analysis provides an opportunity to enhance their relevance by integrating them into a fuzzy-like sociocentric network. This paper was prepared using the Strengthening the Report of Observational Studies in Epidemiology (STROBE) guidelines [33].

A study team member trained in quantitative and network data collection obtained written informed consent from participants and then collected data using interviewer-administered computer-assisted assessments in a private room at the study site. We collected the following data from each participant: sociodemographics, confidant and sexual partner information, mobile phone contact lists, and Facebook friends lists.

Sociodemographic data about the participant and each confidant and sexual partner included first name, last name, gender identity (cisgender man, cisgender woman, transgender or another gender expansive identity), race (Black or not Black), Latine ethnicity (Latine, not Latine), and age (in years). Sociodemographic data were collected using a network inventory using Qualtrics. We downloaded participants’ phone contact information and Facebook network information. We attempted the download of phone contacts and Facebook friends one time, and if the download was not successful then we considered that participants’ data to be missing at random.

We used Cytoscape to visualize participants’ peer referral ties; sexual, social, phone contact, and Facebook networks; and the final fuzzy match networks [47].

The N2 cohort study protocol for the Chicago site was approved by the Biological Sciences Division/University of Chicago Medical Center Institutional Review Board (IRB16-1419) and Columbia University (AAAS7654). Written informed consent was obtained from the study participants. Participants were compensated US $150 for their time and US $20 for each person they referred to the study. A Certificate of Confidentiality from the National Institute of Health was obtained to protect study participants’ privacy.

A total of 412 Black SGM enrolled in the N2 cohort study. Of participants, 387 provided confidant and sexual network data, 273 provided mobile phone contact data, and 144 provided Facebook data. Of these networks, the Facebook network had the highest number of undirected ties (n=1383; undirected means that if one person in the network reported knowing the other, then they were considered to know each other, regardless of the direction of the tie), the fewest connected components, the highest density (13.4% of all potential ties), highest degree (n=19.21), and lowest network diameter (n=5), network radius (n=3), and average characteristic path length (length=2.28). The fuzzy match network consisted of 400 unique nodes and 12 isolates. There was a total of 2054 undirected edges in one centralized component after removing isolates. Its density was 0.026, with an average degree of 10.27. The network diameter was 9, the radius was 5, and the characteristic path length was 3.49. An overview of each of the networks, including nodes, undirected ties, and other network characteristics, can be found in Table 1.

Of 412 participants, 387 completed the social network inventory. Participants named up to 5 social network alters (n=1036; mean 2.73, SD 1.32) and 5 sexual partner alters (n=1017; mean 2.57, SD 1.58). We identified an overlap of 136 confidants who were also sexual partners, for a total of 1917 unique alters.

There were a total of 1917 unique alters. When considering that 412 participants were enrolled in the N2 study, this means that there was a potential of 169,332 directional matches our analyses could identify (ie, A knowing B is treated as separate from B knowing A; between 412 participants, there are 84,666 potentially unique ties as 412 × 411 = 169,332). We found that there were 8020 potential “fuzzy” matches. Upon manual validation, we found that there were 445 undirected ties between the 379 nodes in this network. There were 10 connected components, and the network density was 0.007, indicating it was a sparse network. Average degree was 2.42 and betweenness centrality was 0.028. The network diameter was 19 steps, the radius was 11 steps, and the characteristic path length was 7.77 steps. The egocentric network can be observed in Figure 1. There was no statistical association between betweenness centrality and the HIV care cascade (F₃=0.242; P=.87).

We identified 363 unique ties between 344 participants from the peer referral ties. There were 19 connected components (ie, 19 seeds), and the network density was 0.012. The average degree was 1.99 and betweenness centrality was 0.080. The network diameter was 18 steps, the radius was 9 steps, and the characteristic path length was 7.64 steps. The peer referral tie network is visualized in Figure 2 using a yFiles hierarchic layout to display the chain [48]. There was no statistical association between betweenness centrality and the HIV care cascade (F₃=2.00; P=.11).

Of the participants, 294 provided consent to share their mobile phone contact data. There were 95,831 phone number entries, of which 51,671 were valid phone numbers (ie, in a 10-digit format). We found that there were 416 unique ties among 270 participants from the mobile phone contact data. Figure 3 displays a visualization of the mobile phone contact network. The average degree was 3.18 and betweenness centrality was 0.03. There was no statistical association between betweenness centrality and the HIV care cascade (F₃=1.03; P=.39).

We downloaded the Facebook friends list of 144 participants. Due to refusal and interruptions in the download process, we were not able to download data for 268 participants. There were a total of 18,255 Facebook friendship ties, posts, and comments. Among these, we identified 1383 undirected ties between N2 participants. There was one connected component with a density of 0.134. The average degree was 19.21 and betweenness centrality was 0.009. The network diameter was 5 steps, the radius was 3 steps, and the characteristics path length was 2.28 steps. The Facebook-only network is visualized in Figure 4. There was no statistical association between betweenness centrality and the HIV care cascade (F₃=1.194; P=.31).

Once all networks were combined, we identified 400 participants who were included in the fuzzy-like sociocentric network. There were 2054 unique ties among participants. The density of our fuzzy match network was 0.025, and there was one connected component. The average degree was 10.21 and betweenness centrality was 0.006. The network diameter was 9 steps, the radius was 5 steps, and the characteristic path length was 3.49 steps. The fuzzy match sociocentric network is visualized in Figure 5. There was a statistically significant difference between betweenness centrality and the HIV care cascade (F₃=5.837; P<.001). We conducted a post hoc comparison using the Tukey Honestly Significant Difference test and found that the not virally suppressed group had higher betweenness centrality measures relative to the group not using PrEP (Diff=0.006, 95% CI 0.002 to 0.009; P<.05); the PrEP-using group had lower betweenness centrality measures relative to the not virally suppressed group (Diff=−0.005, 95% CI −0.009 to −0.007; P<.05); and the virally suppressed group had lower betweenness centrality relative to the not virally suppressed group (Diff=−0.005, 95% CI −0.009 to –0.001; P<.01). The other 3 conditions were not statistically significant (P>.05).

We found a low correlation and QAP correlation between HIV status and care engagement between participants and their network members (Table 2). Only the regression model for engagement in care was found to be statistically significant; participants had a 2.07 higher odds ratio (95% CI 1.09‐3.97; P=.03) of engagement in care with higher network-level engagement in care. None of the unadjusted linear network autocorrelation models were significant when examining HIV status and care engagement for each of the different networks (Table 2, models a-e). For the linear network autocorrelation models adjusted for HIV status (Table 2, models f-j), HIV status had a significant negative association with care engagement for all network models: confidant and sexual (β=−0.268, SE 0.048; P<.001), peer referral (β=−0.281, SE 0.049; P<.001), phone contact–only (β=−0.231, SE 0.056; P<.001), Facebook-only (β=−0.252, SE 0.079; P=.001), and the combined sociocentric network (β=−0.256, SE 0.047; P<.001). Results from the linear network autocorrelation model for the combined sociocentric network adjusted for HIV status showed that care engagement had network clustering. Thus, participants who were proximate to network members who were engaged in care were significantly more likely to be engaged in care (ρ=0.128, SE 0.064; P=.045).

Sociocentric network study designs are resource heavy to conduct and primary data collection is difficult. To address this gap, we aimed to develop and test a matching algorithm consisting of multiple types of nonsociocentric networks. We generated a fuzzy matched sociocentric-like network based on the egocentric social networks, egocentric sexual networks, peer referral ties, egocentric mobile phone contact data, and Facebook network present within the N2 cohort study of Black SGM. In addition to analyzing a fuzzy match algorithm, we found that sociocentric networks and digital networks have high utility in unveiling hidden connections. More specifically, we found that in the “fuzzy” network, bridging (ie, betweenness centrality) differed based on the HIV care cascade and engagement in HIV-related care had an autocorrelative effect; these findings were not identified in the partial networks. To our knowledge, this is one of the first studies to harmonize multiple networks to construct a larger fuzzy network, to examine the peer effects of network-level, status-neutral HIV care at the individual level, and to prioritize Black SGM social networks. Our methods and findings are novel and contribute toward the fields of social and digital epidemiology by introducing a fuzzy matching algorithm that can identify hidden ties among participants within different types of networks to better understand the larger sociocentric network.

The initial peer referral chain identified 362 existing ties within our network; however, by including alternative types of networks such as social networks and sexual partner networks, phone contact data, and Facebook friends lists, we were able to identify a total of 2054 undirected ties among 400 participants. The egocentric network quintupled the network degree and doubled density while introducing an additional >50 participants into the network. Whereas the peer referral network had 19 connected components, the final and more complete network brought together these networks into one connected component. In addition, the network diameter decreased from 18 to 9. This means that, while the most distant pairs of nodes required 18 steps at maximum to reach each other, this number halved to 9 steps in the fuzzy match network. In addition, the network radius decreased from 9 to 5, meaning it did not take more than 5 steps for anyone in the network to reach someone else. The characteristic path length, or what Milgram (1967) [49] referred to as “degrees of separation,” also halved from 7.64 to 3.49 steps from the peer referral ties to the fuzzy network. Although the Chicago MSA is indeed a small world, and our population may be too homophilous for a pair of individuals to be selected at random, our findings support the idea of the “small world problem” or the probability that 2 random people selected from a larger population will know each other. Our findings support the need for study designs using peer referral to include additional estimators to account for the interdependence of participants, as these networks may not be as apparent. Further, our study supports the utility of using a fuzzy match algorithm to construct a larger sociocentric network based on other networks, which scientists may collect at a lower burden. Our findings must be interpreted with caution as the network autocorrelation parameter was small. In real-world settings, this may suggest that there is weak social influence within the fuzzy-like network, networks are diverse or mixed and have low homophily, or individual variability may be more important than network influence.

The digital data networks identified the most ties and had superior network density, degree, diameter, radius, and characteristic path length relative to the egocentric confidant and sexual networks and the peer referral network. Interestingly, the characteristic path length of the Facebook network, 2.28 steps, is lower than reported in another study, which found pairs of random people on Facebook to have a characteristic path length of 4 steps [50]. Although our study included mobile phone contact networks and Facebook networks, future studies may have better success in identifying networks using other types of digital networks based on their priority population. For example, studies prioritizing people aged 18‐29 years may have success using Instagram as 78% of people in the 18‐29 year age group use Instagram, which is higher than other age groups (59% of people aged 30‐49 years, 35% of people aged 50‐64 years, and 15% of people aged ≥65 years) [14]. If studies hope to reach Black populations, TikTok or Instagram may be the best social networking sites as 49% of US Black adults report using TikTok (relative to 28% of White adults, 49% of Hispanic adults, and 29% of Asian adults) and 46% report using Instagram (relative to 43% of White adults, 58% of Hispanic adults, and 57% of Asian adults). In addition, after the COVID-19 pandemic stay-at-home orders in the United States, social media platforms became an online “home” for many SGM [51,52]; however, further study on social media use by young (ie, 16‐34 years old) Black SGM is warranted.

Our findings suggest that, while we did not identify differences in betweenness centrality based on a participant’s HIV care cascade in the partial networks, we did so in the fuzzy-like network. More specifically, we found that participants who were virally suppressed had lower betweenness centrality measures relative to individuals in the other HIV care cascades (ie, people who did not use PrEP, people who used PrEP, and people who were not virally suppressed). This means that, relative to individuals who are HIV-negative or not virally suppressed, participants who are virally suppressed appear to be a bridge in the network less often. This could be due to their lower likelihood to engage in behaviors with enhanced vulnerability to HIV such as condomless sex [53], which could also reduce their network connections. People who are virally suppressed may also be embedded in clinical networks instead of other social, sexual, or other community networks, as they are engaged in their HIV-related care. People who are virally suppressed may not only have smaller networks but also have networks that are more closely knit, which could reduce their bridging. Alternatively, people living with HIV may also experience stigma from HIV-negative people, which could reduce their network bridging potential. Additionally, research can explore the relationship between bridging or betweenness centrality and HIV care cascade outcomes.

Our findings must be considered within their context. There is a dearth of studies comparing influence in digital and in-person networks, and it is difficult to ascertain whether endogenous influence and social norms can be transmitted through digital or in-person interactions. Additionally, the weight social media influencers and content creators may have on social norms surrounding health and how social media algorithms “push” specific content through are largely unexplored. Further, we were unable to ascertain which relationships may be more important to intervene on. Finally, within social epidemiology and social network studies, the contextual factors of the sample as well as the geographic setting are important to consider as findings may (or may not) apply to other marginalized populations or geographic settings. For example, Black SGM in Chicago, Illinois, experience intersectional discrimination [20-23] and health disinvestment within their communities [54], which negatively impact health outcomes including HIV-related outcomes. Within this population, social norms may influence access to resources, support, and other opportunities [55-58]. These findings may translate to other Black SGM born in the United States, but may not translate to Black SGM born elsewhere who migrated to the United States or Black SGM living in cities with a more robust health infrastructure. Future studies could explore these topics.

Sociocentric data, the “gold standard” of social network data, are difficult to collect due to resource demands and complex study designs. However, constructing sociocentric networks from digital data may surmount challenges such as limitations in data collection as asking participants to list everyone that they know may be subject to measurement issues like recall bias. Although our inclusion of digital data expanded the breadth of our sociocentric network, the depth, or the influence of individuals in digital networks, is unknown. In addition, although we found that the inclusion of digital data resulted in a model that found statistical significance in HIV-related behaviors, it is unclear if this model is more precise or accurately describes the data. Thus, to see if the quality or strength of digital networks may be similar to that of in-person networks, future studies identifying the quality of these ties are needed. Another way to do so could be to analyze a behavior that is more susceptible to network structure, such as direct interactions among participants (eg, likes, comments, reactions). Although our findings support the collection of additional network data, from an ethical perspective, this may increase the likelihood that a participant may be identified as part of a network that one may not want to be identified with. These concerns must be considered in study designs.

Network analyses are often not perceived to be generalizable because they are representative of one network at a given time point; however, we argue that the purpose of this analysis is not to generalize our findings to the larger population but rather to explore novel fuzzy matching methods to enrich social network research by identifying opportunities to harmonize multiple types of data to triangulate a more robust and complete network. Differentiating selection from peer effects in this analysis may manifest differently based on the mechanisms at play and the networks being considered—for this, Simulation Investigation for Empirical Network Analysis (SIENA) models may allow for the distinguishing between these 2 mechanisms. In addition, participants were asked to provide their perceptions of the sociodemographic information of their network members. Race and ethnicity are complicated social identities that can be incorrectly perceived by others. This could introduce misclassification as participants may incorrectly estimate their network member’s sociodemographic information and introduce bias into our findings by creating incorrect linkages between participants and creating errors in the network structure [9]. This could then trickle down and impact our tests of differences and associations and, if nonrandom, create confounding effects. Because the majority of participants felt close or very close (>80%) to their social members, we believe that misclassification bias may be minimized. A future study could also assess the accuracy of this fuzzy matching analysis. This type of data validation could include a brief survey or interviews with a random sample of the network to authenticate relationships. Another limitation is the number of participants for whom we were able to download complete Facebook data (n=144, 35%). If we were able to include Facebook data from additional participants, we could have potentially constructed a larger fuzzy network. However, because we were able to identify participants who did not provide their Facebook data from the Facebook friends lists of current participants, we believe this bias may be minimized. Future studies can opt to collect less Facebook data, such as only collecting friendship networks, instead of collecting data on group membership, comments, and other Facebook interactions. Finally, we used a cross-sectional linear network autocorrelation, a model that assumes continuous dependent variables, to analyze binary variables. The use of this model assumes that associations are linear, which may not hold for binary variables that constrain probabilities within the (0,1) range. The use of the linear network autocorrelation model may introduce misestimation of effect sizes due to nonlinearity, predictions outside the valid probability range, and heteroskedasticity. To address this bias, we have added supplemental analyses of unadjusted logistic regressions to test the effects of the network exposure to HIV serostatus and engagement in care. To establish peer effects, longitudinal data and a robust set of covariates could be used instead. Our findings may not be capturing peer effects, but rather homophily in the select effect (ie, participants deliberately seek others similar in their HIV care to be friends with). As our study was a pilot approach, our findings on autocorrelative effects may be interpreted with caution as they are marginally significant correlations.

Although not a limitation, a notable feature of this analysis was the time intensity of the match between participants and the social alters and sexual partners listed by other participants. This match of respondents to social network members and sexual network members generated over 8000 potential matches. We manually validated each match to determine if it was a true match or a false-positive match. This was a time-consuming analysis that unveiled <1% of ties of the final network. A machine learning algorithm, if trained appropriately, may alternatively be used to identify matches.

Digital networks, which we operationalized as phone contact and Facebook networks, were the most productive in identifying ties and may be best to characterize social networks. Our experimental approach identified opportunities to characterize networks more thoroughly and accurately. Future studies can explore how to operationalize and examine the influence of ties across various networks. In addition, there are added benefits to including digital networks to egocentric networks as they can identify hidden ties. There are clearly differential impacts of digital data networks on health and HIV outcomes; future studies could more thoroughly assess these networks. Thus, fuzzy matching may be an effective means to identify hidden ties among participants within different types of egocentric and digital networks to better understand the larger sociocentric network, including among vulnerable populations.