The Alter-Ego study is an on-going (2017-) longitudinal network-based study of young (aged 18-30) PWID (egos) and their injection network members (alters). Baseline interview data were used for the analysis.
Eligibility. To be eligible, ego participants (i.e., initial participants who were asked to recruit their network members) had to be (i) 18-30 years old, (ii) injected drugs at least once in past 30 days, (iii) willing to recruit PWID ≥ 18 years old at baseline with whom they injected drugs in the past 6 months (i.e. injection network members); (iv) willing to be tested for HIV and HCV, and (vii) residing in the Chicago Metropolitan Statistical Area in the past 12 months. The injection alters were eligible if they were (i) ≥ 18 years old and (ii) had injected drugs with the ego in the past 6 months. Current IDU was verified by experienced study staff checking for injection stigmata and, if questionable, using a standardized procedure to evaluate participant knowledge of the injection process. Age was verified with a driver’s license or a state ID card. Project staff offered to assist those without identification in obtaining it. Figure 1 shows the sample generation process.
Ego recruitment. The study was conducted at two field sites of a community outreach center located in Chicago, Illinois, USA that has been providing services (e.g., syringe service programs; and HIV and HCV counseling, testing, and case management) and conducting research on people who use illicit substances for over 30 years. The field sites are located in areas that have rates above the city’s average for HIV/AIDS, sexually-transmitted infections, viral hepatitis, and arrests for drug-related offenses and attracts both urban and suburban PWID. We recruited most egos from the syringe services program (SSP) at these field sites. In addition, SSP-recruited participants were screened to ascertain if they obtained syringes at the SSP for other people who reside in the suburbs. Those who did were offered a coupon to refer to the study an age-eligible peer who did not use the SSP or purchase/use drugs in Chicago. To encourage peer-recruited PWID to participate, we used a mobile outreach van staffed with an interviewer/phlebotomist, to conduct data collection off-site near the recruit’s residence or other mutually agreed upon locations. Alternative outreach methods targeting non-SSP suburban PWID included direct recruitment in drug market areas and at community health events using an outreach van, flyers posted at community-based organizations serving PWID, and social media and other online ads. Screening and enrollment of non-SSP PWID from drug market areas were done by indigenous field staff with extensive experience working in these areas and recruiting for similar studies.
Alter recruitment. At their baseline visit, we asked each ego participant to recruit up to five alters in their core injection network, defined as people with whom they injected drugs at least once in the past six months, using recruitment coupons that provided information about the study and were linked to the recruiting ego via alphanumeric code. Coupons could only be redeemed by alters named by an ego participant during their survey. Data collection from alters was required to occur within 6 months of the ego’s baseline visit.
Procedures. All participants (egos and alters) completed a process of informed consent. All study procedures were approved by an Institutional Review Board at the University of Illinois at Chicago. Participants received compensation of $20/hour for the interview. Most participants completed the survey within a 2.5 hour session that included a break (average $50). In addition to hourly compensation for interviews, egos were reimbursed $10 for each alter enrolled.
All participants received HIV and HCV testing and counseling. All services (e.g., SSP; HCV and HIV testing, counseling, and case management; and linkage to medical care) were made freely available to all PWID screened, regardless of study enrollment.
Participants completed a baseline computer-assisted interviewer-administered questionnaire, including background demographics, substance use, HCV testing, injection-related behaviors, and other measures. Egonet data were then collected from participants using GENSI software (24). The survey is touch screen enabled and the participant can tactilely participate in the collection of network data via the “binning” of software-generated nodes which represent the members of their IDU, sex and support networks, with the interviewer. The interview begins with the participant being asked to identify and generate their injection, sexual, and social support networks. They are first asked to generate their total networks (anyone they have injected with, had sex with, or received support from at least once in the past six months). They are then asked to report the gender, age, primary method of drug use and substance, race/ethnicity and county of residence for each. For social support network members, participants are asked what their relationship is to the network member (friend, relative, etc.). Once they have generated the total network, they are asked to identify which of the members are in their core injection networks, core social support networks, and core sex networks (see definitions below). The remainder of the survey focuses on these core networks. Starting with the participants’ IDU core, they are asked questions related to each member and are asked to “bin” the bubbles that represent their network members in the appropriate answer bin. They are also asked how well the members of their 3 networks know each other to collect the strength of tie for each relationship.
Injection network. Participants were asked to identify people who they used drugs with in the past six months (total injection network). Among those, they were then asked to identify a core injection network of up to ten people who they used drugs with more than once in the past six months. Data on the injection network were collected by proxy for the total and core injection network, and directly from up to five alters identified and recruited by egos.
Support network. Participants were asked to identify people who provided support to them in the past six months (total support network). This included “anyone you could talk to about things that are personal and private or get advice from if a situation came up” (personal support), and “anyone that would let you stay at their place if needed” (shelter support). Among their support network, they were then asked to identify a core support network of up to ten people who provided the most support in the past six months.
Sex network. Participants were asked to identify people they had sex with in the past six months, including vaginal, anal, or oral sex. Among these, they were then asked to identify a core sex network of up to ten people who the participant had sex with more than once in the past six months.
Individual measures (ego and alters)
Demographics. Self-reported gender, age, race, and Hispanic/Latinx ethnicity were collected. Gender was reported as male, female, transgender, or other. Multiple categories could be selected for race; ethnicity was indicated separately as Hispanic/Latinx or not Hispanic/Latinx. Race and ethnicity were also combined to create an indicator variable with mutually exclusive categories: non-Hispanic white (white only), non-Hispanic Black (including mixed), Hispanic (any race), and non-Hispanic other race. Employment was assessed based on responses to the question “During the last 6 months, did you receive any money from any of these sources?” Participants who indicated that they received money from a regular job (full or part-time) or self-employment were classified as employed. Egos were similarly asked to provide sociodemographic information on their alters as part of the network interview.
HIV and HCV status. Rapid HIV and HCV antibody (Orasure Technologies, xx) testing was used to determine prevalence of exposure (past or current infection) for egos and recruited alters.
Syringe and equipment sharing. Injection behavior was assessed on a 5-point Likert-type scale (never, less than half the time, about half the time, more than half the time, or always). Questions assessed receptive syringe sharing (RSS) (“When you shot-up in the last 6 months, how often did you use a syringe that you know for sure had been used before by someone else?”), equipment sharing (“When you injected drugs in the last six months, how often did you use any of the following items with other people? (a) drawn from the same cooker, (b) used the same cotton, (c) used the same rinse water”), and backloading (“In the last six months, how often did you inject with a syringe AFTER someone else has squirted drugs into it from their syringe?”) Equipment sharing was coded as the maximum of the three items (cookers, cottons, water). Each of these measures was dichotomized to indicate any behavior in the past six months.
Summary measures of ego network size, composition and structure were computed for the IDU, sex, and support networks, which are defined and described in an additional table [see Table S1 in Additional File 1]. Ego-alter tie strength was based on frequency of contact, assessed by asking “How often do you talk to or see this person?” with options on a 6-point scale from every day to less than once a year. The last 3 categories were collapsed into once a month or less to create a 4-point scale. Alter-alter tie strength was based on the ego’s assessment of how well the two people knew each other, on a 4-point scale with anchors 1 = “casual acquaintance” and 4 = “very close relationship.”
All network measures were computed using NetworkX version 2.4,(25) except modularity, which was computed using the igraph R package version 1.2.5 (26). Modularity was computed by finding the network partition that maximizes Newman’s Q. Hierarchy was implemented following Eq 2.9 of Burt.(27) For measures that rely on distance rather than tie strength (for example, betweenness and closeness centrality), the scale was reversed, so that strongest ties were coded with a distance of 1 and weakest ties with a distance of 4. Average and local clustering, average distance, centrality, centralization, constraint, effective size, efficiency, and hierarchy were all computed both with real-valued edge weights based on tie strength and with binary edge weights based on presence/absence of a tie. Centralization measures and hierarchy are undefined when ego has degree one; these were recoded to zero for analysis.
Imputation of missing ties. Missing ties and tie strengths were imputed using sklearn version 0.22.1.(28) For ego-alter tie strength, a random forest classifier was trained using all available ego-alter tie strengths, with ego sex, alter sex, and relationship types as features. Tie strength was randomly sampled for missing data according to the predicted probability distribution from the classifier. For alter-alter ties and tie strengths, a random forest classifier was trained using all alter-alter networks without missing data, with the sex of each alter and the relationship types between ego and each alter as features. Tie presence and strength was randomly sampled according to the predicted probability distribution from the classifier.
Covariates and Outcomes. Covariates included 40 injection network variables, 40 support network variables, and 2 sex network variables, plus ego demographic variables: age, sex (female vs. male), and binary indicators of white, Black, and Hispanic. Risk behavior outcomes were binary indicators of RSS, equipment sharing, and backloading in the past six months.
Sample. Figure 1 summarizes the unique participants (egos and alters). Of the 177 egos who completed the baseline study visit and provided proxy data on their total and core injection, sexual and support networks, 9 were excluded: 6 did not have any support network and 3 had missing data on key variables. Of the 117 alters who were recruited into the study and completed the baseline study visit in person, 16 were excluded: 9 had not injected drugs in the previous 6 months, 6 did not have a support network, and 1 had an incomplete network interview. In addition to providing data on relationship with egos, recruited alters also became second wave egos. The final analytic sample includes 168 ego and 101 alter (second wave ego) participants (total=269).
Analysis. Given the large number of network measures potentially associated with risk behavior, some of which are highly correlated with one another, we conducted adaptive lasso logistic regression (29, 30). The advantage of lasso regression is the k-fold cross-validation, which estimates prediction error by resampling to select potentially important predictors of HCV antibody positive status and injection risk behavior from a large set of network variables. Lasso regression is a supervised machine-learning method in which the goal is to obtain the subset of predictors that minimizes prediction error. The lasso procedure performs both variable selection and regularization, using shrinkage to select a subset of predictors, which prevents overfitting and encourages simple sparse models (31, 32). Analyses were conducted using Stata (v. 16) .
Adaptive selection in Stata uses cross-validation (CV) to select lambda (the shrinkage parameter), but multiple lassos are performed. In the first lasso, a lambda is selected, and penalty weights are constructed from the coefficient estimates. Then, these weights are used in a second lasso where another lambda is selected. Variables with zero coefficients are discarded after each successive lasso, and variables with nonzero coefficients are given penalty weights designed to drive small coefficient estimates to zero in the next step. CV is done by dividing the data randomly into folds. One fold is chosen, and then a regression is fit on the other folds using the variables in the model for that lambda. With these new coefficient estimates, a prediction is computed for the data of the chosen fold. The mean squared error (MSE) of the prediction is computed. This process is repeated for the other folds. The MSEs from all folds are then averaged to give the value of the CV function. We repeated the analysis with 4 different random seeds and increased the number of folds until a result was replicated consistently.
HCV and risk behavior outcomes (RSS, equipment sharing, and backloading) were regressed on the 40 injection network variables, 40 support network variables, 2 sex network variables, ego age, sex (female vs. male or undetermined), and binary indicators of white, Black, and Hispanic.
Exploratory analysis. We conducted post hoc logistic regression analyses and computed predicted probabilities to explore the relationships between selected predictors and outcomes.
Sensitivity analysis. We conducted sensitivity analyses to compare the results using imputation with alternative strategies for missing data, including (1) listwise deletion, and (2) imputation of ego-alter tie strengths only, treating missing alter-alter ties as non-existent.