Participants complete HIV/STI testing after providing consent and completing the baseline assessment and the 12-month follow-up assessment. Specifically, a testing counselor affiliated with the study and based at a Romanian non-governmental organization specialized in HIV/STI prevention and treatment for GBM and other marginalized populations contacts the participant to provide two options: 1) mail an HIV/STI test kit to their home or 2) have the participant pick up the test at their offices. Participants receive a self-testing kit containing a rapid HIV/syphilis test, and swabs and urine collection container for pharyngeal, rectal, and urethral testing of chlamydia and gonorrhea. The testing counselor guides participants through the testing steps [82, 83]. The HIV and syphilis test results are available within 20 minutes of testing. Participants are required to take a photograph of the results (marked with their study ID) and upload it to a study platform. The chlamydia and gonorrhea self-collected samples are mailed back to the organization, which mails them for analysis to a laboratory. Upon receiving results for chlamydia and gonorrhea testing (usually within one week after lab receipt), the testing counselor communicates the results to the participant. For positive or inconclusive test results, the testing counselor provides participants with the name and contact information of the study-affiliated infectious disease provider in their area for confirmatory testing and treatment, as appropriate, and offers to assist with this linkage. Finally, participants were asked to complete a satisfaction survey at their 12-month follow-up [84].
Data Analyses
Primary and Secondary Outcomes
It is hypothesized that participants randomized to the Comunică intervention, compared to those randomized to the EAC, will report significantly greater decrease on the primary outcome (HIV-risk-transmission behavior defined as condomless anal sex acts with HIV-positive or unknown-status partners outside of the context of one’s own adherent PrEP use or primary partner’s adherent PrEP use or undetectable viral load in the past 30 days) and secondary outcomes (depressive and anxiety symptoms, suicidality, heavy alcohol use, and HIV/STI testing outside of study testing) outcomes at 4-, 8-, and 12-month follow-up visits.
Sample Size Justification
As we approached the end of study recruitment, three factors transpired for estimation of power: i) While 305 participants were finally enrolled, we had estimated that we would likely only be able to recruit 288 subjects by 31 December 2022 (our date of termination of recruitment) and we therefore re-estimated power based on this number; ii) The overall participation rate for attending all three post-intervention visits was 86% (with a high proportion of the remaining sample completing two post-interventions assessments); iii) We had normative distributions of the study outcomes that could be used in establishing a study design and metrics for powering the primary outcome. To that end, we re-estimated the study power as described below. This new estimate was approved by the Data Safety and Monitoring Board (DSMB) and our program officer at the National Institute of Mental Health.
In the original proposal, a repeated measures mixed linear model (or generalized estimation equation) using all four study visits (baseline, and 4-, 8-, and 12-month follow-ups) was planned. However, the distribution of our primary outcomes (HIV-transmission-risk behavior) based on the data we now had near the end of the study was problematic for fitting this type of model, which is based on a central limit theorem assumption that may not manifest. Namely, there were two issues to consider: 1) at each follow up visit, over 50% of participants manifest no HIV-transmission-risk behavior creating a large point mass at the lower limit of 0. 2) On the other end, the distribution of HIV-transmission-risk behavior acts was very skewed with numbers of acts greater than 50 leading to skewness > 3, which again contradicts normality-based methods.
Although generally accepted approaches for power estimates for this situation do not readily exist, we found one approach that will satisfy assumptions for standard normal methods and that is also amenable to conducting a power estimation. That approach is to take the average behavior over all three follow-up timepoints (4-, 8-, and 12-month assessments) as a single within-person outcome, rather than evaluation of the repeated measures at 4-, 8-, and 12-month follow-ups as separate within-person outcomes. For example, if a person reported 0, 1, and 2 high risk acts respectively at the 4-, 8-, and 12-month follow-up visits, these would be summed together (0 + 1 + 2) = 3 and averaged over the 3 visits (3/3 = 1) as a single outcome of one high risk sex act over the previous 30 days per follow-up period. The details on what happened when we examined this outcome in our (still unblinded to intervention) data follow in the next paragraphs.
When we took the average of all three follow-up timepoints of HIV-transmission-risk behavior acts (i.e., during the previous four months), only 24% of participants had no HIV-transmission-risk behavior acts over the entire 12-month time period (i.e., the point mass at 0), which given our projected sample size of 288 (or 246 evaluable assuming 14% loss of data as we have been observing) is low enough to treat this average as a continuous variable in a normal approximation. However, due to a few individual(s) reporting very high numbers of risk acts (i.e., > 50), this variable was still skewed (> 3) so we capped (i.e., Winsorized) the maximum number of average per-assessment HIV-transmission-risk behavior acts at 15 acts, the upper 97th percentile of HIV-transmission risk behavior. When this was done, the HIV-transmission-risk behavior outcome was close enough to normal (i.e., skewness = 1.5) for the central limit theorem to apply to the linear model presented below.
The power estimation approach assumes that a linear model will be fit Y = a + b X + c T + ε
Where
Y = Averaged (during prior 4 months) prior 30 day high-risk sex behavior over 12-month behavior
a = intercept
X = baseline prior 30-day HIV-transmission-risk behavior acts
T = 0 for control, 1 for intervention
ε is random error with mean 0 and constant variance
And a, b, c are unknown parameters that are estimated in the model fit
The null hypotheses c = 0 will be tested with an overall two-sided type 1 error of 0.05.
Importantly for power estimation, the correlation of baseline and the average 12-month HIV-transmission-risk behavior acts was 0.37, which means that after adjustment for the pre-intervention behavior, the standard deviation of the post-intervention behavior would be the square root of (1-0.372) = 0.93 that of the unadjusted outcome. Based on this and with 288 subjects, 86% of whom participated in all 3 visits, as we have been observing will happen (or conservatively 123 in each treatment arm), there is 80% power to detect an effect size of 0.33, slightly greater than our originally estimated effect size of 0.25–0.27 and at the upper end of the range of effect sizes found for behavioral interventions addressing multiple health outcomes among sexual minority men [85]. We believe that 0.33 represents a plausible effect size to detect in this trial given the strong distinction between the two intervention conditions, with one involving an active therapist-guided intervention and the other consisting of self-guided psychoeducation only. The standard deviation of the (Winsorized) averaged outcome over 12 months per-assessment HIV-transmission-risk behavior acts was 3.46 acts. Multiplying this standard deviation by the effect size of 0.33 gives 1.14 HIV-transmission-risk behavior acts. This means that the study will have 80% power to detect an overall mean reduction of 1.14 HIV-transmission-risk behavior acts in the prior 30 days per assessment period in the intervention compared to the control condition.
It should be noted that our final analysis will most likely incorporate the partial information from men with only 1 and 2 post-intervention follow-up visits through imputations, or more exactly adjustment of the partial information for number of follow-up periods reported, If so, this would increase power, albeit by a very modest amount.
Data preparation
Skewed variables will be recoded for analytic symmetry using appropriate log, square root, or other non-linear transformations. Should we fail to be able to find a transformation that achieves sufficient linearized normality, then robust generalized estimation equations (i.e., with logit or log link) will be fitted to dichotomized or count outcomes. We will also examine variable distributions, which may suggest dichotomous and multinomial recoding relevant to our primary research questions and would increase the statistical power of our models. Dependent and independent variable values will be cross-plotted as a function of time-in-study and summarized using parametric and non-parametric modeling methods such as loess curves. In addition to detecting trends and temporal patterns, graphic representations of time series data will provide knowledge of within- and between-individual variability of measurements. Results will be used to construct more complex cross-group/-time analytic models (below).
Analytic Plan
The analysis of the primary and secondary outcomes will use intent-to-treat with participants analyzed according to their original treatment assignment. SAS 9.4, SPSS 26.0, Stata and/or R software will be used for all analyses.
Comparability of Treatment Groups
Differences in baseline demographic characteristics between the two treatment arms will be assessed using appropriate graphical and statistical methods including summary statistics and p-values from exact, rank, chi-square, t-tests and ANOVA. Of note, as analyses progress, we will control for variables related to the study outcome in the analyses. We will also investigate if the randomization scheme was compromised.
For the HIV-transmission-risk behavior primary outcome (number of condomless anal sex acts in the past 30 days with HIV-positive or unknown-status partners outside of the context of one’s own adherent PrEP use or primary partner’s adherent PrEP use or undetectable viral load) analyses, the statistical significance threshold for an intervention (vs. control) arm effect will be a two-sided p ≤ 0.05. The primary outcome will be evaluated between the two arms at 4-, 8-, and 12-month follow-ups combined in a repeated measures analysis that adjusts for baseline behavior. This will be analyzed as using negative binomial regression with baseline, and 4-, 8-, and 12-months clustered within the same person. Main effect terms for 4-, 8-, and 12-months post-baseline (each timepoint vs. baseline) will be included in the model. A single interaction term between the 4-, 8-, and 12-month measures with the intervention arm will be included in the model to test for pooled post-baseline treatment arm differences. The relative number of HIV-transmission-risk behavior acts with a 95% confidence interval about his term will quantify intervention effect. Generalized estimating equations (GEE) with person as the cluster will be used to account for within-person repeated measure collinearity. As a sensitivity analysis, this will be repeated including all baseline covariates that are statistically associated (p < 0.05 to enter and p ≥ 0.10 to leave in a stepwise selection) with HIV-transmission-risk behavior in negative binomial GEE models with person as the cluster and adjusting for time of visit (e.g., 4-, 8-, and 12-months each vs. baseline). If there are excess zeros at each post-baseline visit, we will consider using a zero-inflated negative binomial model instead. However, this approach will split the intervention effect parameter into two models and thus may dampen power to detect statistical significance for an intervention that affects both parts. In this setting, we thus will more likely use the sensitivity analysis approach described below.
As a further sensitivity analysis, averaged HIV-transmission-risk behavior over all three follow-up visits (or two post-baseline visits if one visit is missing) will be used as the outcome in an ANCOVA linear regression model that adjusts for baseline HIV-transmission-risk behavior as a predictor and includes treatment arm assignment as a covariate. For those who are missing one follow-up visit, indicator variables as to which visit is missing will be included. The mean difference in HIV-transmission-risk behavior acts with a 95% confidence interval about his term will quantify intervention effect. This will be repeated including all baseline covariates that are statistically associated (p < 0.05) with HIV-transmission-risk behavior in stepwise selection (p < 0.05 to enter and p ≥ 0.10 to leave) into the above model. Should the negative binomial model described above fail to converge, this will become the primary analysis. The statistical significance threshold for the new (vs. control) intervention arm effect will again be a two-sided p ≤ 0.05.
The secondary outcomes of interest (all of these assessed at baseline and at 4, 8, and 12-month follow-up visits) are depression and anxiety symptoms, suicidality, and heavy alcohol use. Depression, as measured by CES-D, will be examined as a continuous variable and as a binary variable using the cutoff of ≥ 16 (indicating clinical depression). Anxiety, as measured by BAI, will be examined as a continuous variable and using the cutoff of ≥ 16 (indicating potentially concerning levels of anxiety). Suicidality, as measured SIDAS, will be examined as a continuous variable, as a binary outcome using the cutoff of ≥ 21 (indicating high risk of suicidality), as well as any score above 0. Heavy alcohol use, as measured by AUDIT-C, will be examined as a continuous variable and as a binary outcome using the cutoff of ≥ 4. Percentage of heavy drinking days in the past 30 days prior to the visit, as measured by the TLFB, will be examined as a continuous variable. Due to multiple comparison issues, these will each be tested individually using a two-sided Type-1 error of 0.01 and quantified using 99% confidence intervals. Levels of these measures at 4-, 8-, and 12-month follow-ups will be compared (adjusting for the level at the baseline visit) between the Comunică intervention and EAC in repeated measures analyses as described below.
For continuous outcomes that are heavily skewed to the right and without an excessive point mass at 0 for 4-, 8-, and 12-month follow-ups (e.g., the BAI or number of heavy drinking days), a similar approach to that described for the primary outcome analyses will be used. For continuous outcomes that are not heavily skewed to the right at 4-, 8-, and 12-month follow-ups (e.g., CES-D or AUDIT-C) repeated measure linear regression mixed models will be fitted for outcomes at baseline and 4-, 8-, and 12-month follow-ups with subject intercept as a fixed effect, main effects for 4-, 8-, and 12-month follow-ups, and a single interaction term between treatment arm assignment and the timepoint being follow-up. In sensitivity analyses, this will be repeated including all baseline covariates that are statistically associated (p < 0.05) with the outcome in models using stepwise selection (p < 0.05 to enter p ≥ 0.10 to leave). The mean post-intervention difference in the outcome between the treatment arms with 99% confidence intervals about this term will quantify intervention effect. Binary outcomes (e.g., HIV-transmission-risk behavior > 0, CES-D ≥ 16, SIDAS ≥ 21, SIDAS ≥ 0, AUDIT-C ≥ 4) will be analyzed in repeated measures models using the baseline visit and follow-up visits at 4-, 8-, and 12-month follow-ups. Repeated measures generalized estimation equations will be fit about individual as the cluster using a logit link. The visit number (i.e., 4-, 8-, and 12 months vs. vs. baseline) and treatment arm assignment will be included as main effects, as will the baseline level of the outcome being modeled. In sensitivity analyses, this will be repeated including all baseline covariates that are statistically associated (p < 0.05) with the outcome in models using stepwise selection (p < 0.05 to enter p ≥ 0.10 to leave). The intervention effect will be quantified by odds ratios with 99% confidence intervals. Finally, we will also use exact tests to compare treatment arms for having been ever diagnosed during the single timepoint 12-month study follow up with HIV, syphilis, chlamydia, and gonorrhea.
Mediation analyses
In our mediation analyses, we will examine whether changes in the proposed mediators (e.g., self-efficacy for condom use or heavy alcohol use prevention, identity concealment,
internalized homophobia) precede and statistically mediate intervention effects consistent with our IMB and minority stress models. We will use path analysis/structural equation modeling to model and assess the size of the indirect effect from intervention condition to 12-month outcomes through mediators assessed at 4- and 8-months controlling for baseline effects of these mediators.