This study presents the per-protocol effect of zinc supplementation in the ZINC trial under the ideal conditions of high adherence and no loss to follow-up. We used the inverse probability weighting to adjust for baseline and post-randomization confounders. Our per-protocol effect estimate of the mean difference in change in VACS index between the zinc and placebo group was − 8.01 (95% CI -16.42, 0.01), somewhat larger than the intention-to-treat effect difference in change (-4.68 (-9.62, 0.25)), but also not statistically significant. This suggests that the intention-to-treat estimate may have underestimated the beneficial effect of zinc supplementation in HIV-positive heavy drinkers. However, like in the intention-to-treat effect estimate, confidence intervals were wide.
We found that the mean difference in change between individuals in the zinc and placebo group under high adherence was − 4.07, -8.01 and − 12.34 for high adherence defined as 70%, 80% and 90% of pill-coverage, respectively. This apparent trend of increasing effect magnitude with increasing thresholds for definition of high adherence is suggestive of a potential beneficial effect of high adherence to zinc supplementation and requires further exploration through new studies.
The intention-to-treat approach is the gold standard for estimating the causal effect of interventions in randomized controlled trials. However, in double-blinded randomized trials with suboptimal adherence to the study interventions, the intention to treat effect might provide biased estimates and should be complemented with per-protocol effect estimates. Unlike the intention-to-treat effect, the per-protocol effect quantifies the maximum benefit of a treatment/intervention and, usually, clinicians and patients making treatment decisions find the per-protocol effect estimates more informative than the intention-to-treat estimates. G-methods have been typically developed and used for the analysis of large observational studies and for the per-protocol analysis of large randomized trials[9, 11]. Our findings show that approaches based on the g-methods can be used for the per-protocol analysis of relatively small clinical trials like the ZINC trial to complement the intention-to-treat effect estimates.
Our per-protocol effect estimate was similar to the standard per-protocol estimates, which restricted analyses to individuals who reported high adherence over three visits. This similarity should not be interpreted as an argument in favor of standard per-protocol analysis, because such estimates are generally subject to bias. They rely on the unrealistic assumptions that adherence and loss to follow-up occur completely at random. In contrast to the standard approach, we adjusted our analyses for baseline and post-baseline characteristics likely to influence adherence.
Our analysis has limitations. First, self-reported adherence might be an imperfect measurement of pill uptake. A sensitivity analysis using an alternative measure of adherence based on the riboflavin urine test showed a somewhat smaller benefit of zinc supplementation. This discrepancy may arise from the difference in timing of the two adherence assessments: self-reported adherence was defined as the percentage of pills taken in the previous 6 weeks on a visual analog scale, while the riboflavin test allows detection of medication intake occurring in the previous 24 hours. Second, our analyses assume that all prognostic factors that predict adherence are identified and accurately measured. Our estimates would be biased if one or more important determinant of adherence were not included in the model. However, we found no adherence effect in the placebo group. This is reassuring as it indicates that the available data are sufficient to adjust for confounding and selection bias due to loss to follow-up. Finally, because of the relatively small sample size and the large proportion of individuals who were lost to follow-up, the confidence intervals for our estimates were wide. Fully parametric methods, such as the parametric g-formula, may offer an alternative way to estimate the per-protocol effect in a small trial.
In conclusion, we used the rich data collected by the ZINC trial to estimate the per-protocol effect of zinc supplementation on the VACS index at 18 months while adjusting for both non-adherence and loss to follow-up. High adherence to zinc was associated with a lower VACS index score, but the confidence intervals were wide. Our per-protocol effect estimation confirms the potential benefits of daily zinc supplementation in HIV-positive heavy drinkers. Further studies examining larger sample sizes are needed to shed more light on the potential benefits of zinc supplementation in HIV-positive heavy drinkers.