Cluster-randomized trials (C-RCTs) are expensive to conduct. Using surveillance data on new HIV diagnoses instead of measuring incidence in the trial could reduce costs. We used mathematical models to evaluate when surveillance data can be used to estimate impact in HIV intervention C-RCTs.
We used a model of HIV transmission among men who have sex with men in Baltimore, US, to simulate C-RCTs scaling up antiretroviral therapy (ART), pre-exposure prophylaxis (PrEP) and HIV testing in combination or alone. We tested whether modelled reductions in total cumulative HIV diagnoses predict model cumulative HIV incidence reduction over a 2-year trial. We also tested if reductions in diagnoses predict incidence reduction better over a longer trial duration (≤4 yrs) or when measured in later trial years. We explored if reductions in other surveillance measures – diagnoses with acute infection, diagnoses with early (CD4>500 cells/µl) infection, or diagnoses adjusted for testing volume – better predict incidence reduction. We used Pearson correlation to assess precision and report bias and sensitivity to detect a true incidence reduction.
Over a 2-year trial expanding ART+PrEP+testing, model results suggest total diagnosis reductions correlate poorly with incidence reduction(r=0.386), underestimate incidence reduction (by 97%), and have 52% sensitivity (Table). Precision and sensitivity were better in trials expanding ART(r=0.878; sens 100%) or PrEP(r=0.960; sens 88%) alone, but bias remained (-52% for ART, -55% for PrEP). In trials expanding testing alone, diagnoses increased with decreasing incidence(r=-0.915). Measuring impact in longer trials or over later years improved correlations between diagnosis and incidence reductions for trials expanding ART+PrEP+testing, up to r=0.795 over the 4th year, and reduced bias. For ART+PrEP+testing trials, reductions in acute, early or adjusted diagnoses correlate poorly(r<0.51) with incidence reduction. Reductions in acute or early diagnoses correlate sufficiently with incidence reduction only when ART alone is expanded(r=0.993, r=0.953, respectively), but are biased (-18%, -41%).
Modelling results suggest that surveillance diagnoses data can only rarely be used to estimate C-RCT HIV incidence reductions. Reductions in acute/early or total diagnoses may be adequate predictors in C-RCTs expanding ART or PrEP alone if bias can be adjusted for. None of the diagnoses markers explored were appropriate for C-RCTs expanding HIV testing.