JMIR Publications

ABSTRACT

Background:

Young Black and Latino men who have sex with men and transgender women who have sex with men (YBLMT) face disproportionate HIV-related health disparities, with barriers to care exacerbated by intersectional stigma and geographic constraints. Online health interventions offer accessible, stigma-reducing alternatives for HIV prevention and support. However, recruitment strategies influence not only participant demographics but also retention, an important factor in intervention efficacy.

Objective:

This secondary analysis examines how recruitment sources influenced study retention in HealthMpowerment (HMP) 2.0, an app-based randomized controlled trial (RCT). We assessed retention differences across recruitment sources, explored demographic variation by source, and identified predictors of retention using predictive logistic regression modeling.

Methods:

Participants (N=750) aged 15–29 were recruited through four primary sources: ads on social media, ads on partner-seeking apps, study contacts (prior research participants), and other methods (e.g., community outreach). Retention was defined as completing at least three of four follow-up surveys over 12 months. Chi-square tests assessed retention differences across recruitment sources. A machine learning-based predictive logistic regression model, adjusted for class imbalance using Synthetic Minority Oversampling Technique (SMOTE), explored demographic and recruitment predictors of retention.

Results:

Recruitment source was significantly associated with retention (χ²(3)=30.40, P<.001). Participants recruited via social media (90.6%) and study contacts (87.3%) exhibited higher retention than those recruited from partner-seeking apps (73.1%) and other sources (72.3%). Participants recruited through social media were younger, more likely to be Hispanic, and predominantly HIV-negative, while those recruited from partner-seeking apps were older, more racially diverse, and more likely to be HIV-positive (P<.001). Logistic regression confirmed that recruitment source was the strongest predictor of retention, followed by age, race, gender identity, and HIV status. The SMOTE-adjusted model improved recall for low-retention participants but reduced overall accuracy.

Conclusions:

In a national online HIV intervention, participant retention and demographics varied by recruitment source. While social media and study contacts yielded higher retention, partner-seeking apps reached populations at increased HIV risk. Enhanced retention strategies could be tailored and targeted to ameliorate differential attrition. Future interventions should optimize recruitment approaches by balancing cost-effectiveness, retention, and sample diversity to improve retention among YBLMT populations. Clinical Trial: ClinicalTrials.gov NCT03678181