Analysis plan

The primary analysis will be to compare the prevalence of community–level outcomes across intervention and control clusters, adjusting for pre-intervention prevalence levels. For a given outcome, we will first log-transform the 22 cluster-level proportions obtained from the post-intervention RDS. Then, via weighted least squares linear regression, these will be regressed on a dummy term for the control arm (vs. intervention), another for the MSM stratum (vs. PWID), and a term for the log-transformed cluster-level pre-intervention proportions. The exponentiated coefficient for the control arm term is thus the prevalence risk ratio (PRR), and (1-PRR) × 100 % is the percentage increase in service utilization associated with the intervention. The primary analyses will be conducted using the RDS-II weighted cluster-level proportions from both pre- and post-intervention RDS samples. These RDS-II weights, which account for personal network size (number of PWID or MSM seen in the past 30 days), will be calculated using the RDS Analyst Software Version 0.5 (http://hpmrg.org). For secondary outcomes that are continuous (e.g., community viral load) we will use a similar regression approach.

We will conduct several sensitivity analyses for the primary and all secondary outcomes First, we will repeat analyses using unweighted cluster-level proportions from both the pre-and post intervention RDS surveys. Second, we will consider adjustment for demographic covariates (age, sex, marital status and educational attainment) measured at the post-intervention RDS that are associated with the outcome and are differentially distributed across intervention and control clusters. We will consider adjustment for these factors if the p values for associations with the outcome and the intervention vs. control clusters are <0.05 and the OR is >2 or < 0.5. A two-stage approach will be used when adjusting for individual-level covariates: at the first stage, for a prevalence outcome, individual responses are modeled with a logistic regression model adjusting for all relevant covariates except the dummy term for control vs. intervention. In the second stage, observed and expected prevalence counts for each cluster are calculated, followed by t-test-like analyses of log-transformed ratios of observed to expected [53]. We will also consider an approach that models the difference in outcome prevalence between the pre- and post- intervention surveys using the same cluster-level comparison approach. We will also consider individual-level analyses using multi-level random effects regression approaches (Stata GLLAMMs program) to account for dependence of responses within clusters [54–56]. These models allow inclusion of fixed effects (e.g., intervention), random effects (e.g., clusters), adjustment for pre-intervention covariates at the individual and cluster level, and incorporation of scaled RDS weights as sampling weights. Finally, we will conduct descriptive analyses of the HIV care continuum, before and after the intervention phase of the study, in which completion of earlier steps are assumed to be necessary to complete later steps. Additionally, we will consider sensitivity analyses of outcomes in the HIV care continuum, where biologic markers such as HIV RNA and serum antiretroviral drug testing, are used to supplement self-reported data on access to care.

Several subgroup analyses are also planned. First, we will analyze all outcomes separately within each stratum (PWID and MSM). Using the combined sample of MSM and PWID sites, we will further compare all outcomes within subgroups defined by age, marital status, educational attainment, substance use (drug and alcohol use), and personal network size (number of persons in risk group [PWID or MSM] known and seen in the prior 30 days). Using only the PWID sites, we will also analyze subgroup differences by age, sex, marital status, educational attainment, substance use (including alcohol use), personal network size and region. In the MSM sites, we will also analyze subgroup differences by age, sexual identity, marital status, educational attainment, substance use (including alcohol use), personal network size and region. We also plan subgroup analyses by HIV serostatus and awareness of status for risk behaviors, HIV testing of spouses, substance use, stigma, and depression.

Network effects will be ascertained by comparing utilization of ICCs and services within ICCs across networks as defined by RDS. For example, we will examine utilization patterns across recruiters and recruits in the evaluation RDS. We will also ascertain whether utilization of ICCs varied by wave of RDS. Individual-level comparisons will draw on data from post-intervention RDS participants in the intervention clusters, in which extra questions will address participants’ use of ICC services. In addition, biometric data at the post-intervntion RDS will be linked with the biometric data from the ICCs to determine utilization. Individual analyses will use descriptive statistics and log-binomial regression to compare the level of each outcome (e.g., proportion accessing HCT in prior 12 months) by the main exposure of interest (visiting an ICC). We will adjust for individual-level confounders including demographic characteristics. Analyses will use multi-level random effects regression approaches to account for dependence of responses within personal networks [56]. In addition, using the biometric data to link persons between the pre-and post intervention RDS samples, we will conduct exploratory within-individual comparisons of the primary outcome and secondary outcomes. For example, restricting the sample to persons who participated in both the pre- and post intervention RDS samples and are eligible for the outcome, we will compare across control and intervention clusters the proportions of person who transition from not having the outcome to having it and vice versa.