IMPROVING INFERENCES FROM RANDOMIZED TRIALS: USING PER-PROTOCOL ANALYSES TO ESTIMATE EFFECTS OF HIV TREATMENT

Per-protocol effect estimates quantify the effect of treatment on the outcome of interest had participants, potentially counter to fact, adhered to the assigned treatment regimen. When deviations from the trial protocol arise, per protocol effect estimates are informative. We aimed to characterize per-protocol analyses in modern clinical trials and evaluate per protocol effects under different definitions of the treatment protocol.We conducted a bibliometric methods review of trials in five top medical journals between 2020-2025. Outcomes of interest were the proportion and description of published per-protocol analyses, and reporting of select CONSORT 2010 items. Of the 114 studies included, 45 (40%) reported a per-protocol analysis. The proportion of per-protocol analyses were higher in publicly funded trials (66%) versus those funded solely by industry (10%), but did not vary by year. Per-protocol analyses were frequently conducted by excluding those that deviated from the protocol with only 1 study accounting for predictors of deviation from treatment.Next, in the ACTG A5202 trial we evaluated the difference in risk of viral failure or death between TDF/FTC and ABC/3TC at 48 and 96 weeks under intent-to-treat and multiple protocols defined by allowing up to and including 1, 2, 3, and 10 missed doses of the assigned treatment. Inverse probability weights (IPW) that included both baseline characteristics and time varying markers of disease were used to account for informative censoring due to loss to follow-up and missed doses. Of the 1857 ACTG A5202 participants, 338 (18%) reported missing at least one dose and 51 (3%) reported missing 10 or more doses and were more likely to miss a dose if assigned to the ABC/3TC arm. At 48 and 96 weeks the risk differences comparing TDF/FTC to ABC/3TC were similar to the intent-to-treat estimates across protocols. These results suggest there is room for more per protocol analyses to be conducted in trials, and particularly for per protocol analyses that account for possible selection bias. While we didn’t find a difference from the intent-to-treat when altering the protocol definition in the ACTG A5202 study, our approach can guide similar analyses in future trials.