The dramatic fall in HIV and hepatitis C virus (HCV) incidence observed in Amsterdam has been claimed as a success of harm reduction , contributing to, and strengthening, the existing imperfect ecological evidence suggesting that sites with more intervention coverage have lower infection rates [2, 3]. It is not the first time that a decline in disease transmission has been re-evaluated. Other researchers have highlighted the importance of testing (and separating out) whether declines in HIV prevalence are due to natural progression of the epidemic or the effect of an intervention [4, 5]. It has been observed that the decline in HIV among people who inject drugs (PWIDs) in Spain occurred after harm reduction was scaled up . Famously, it has also been debated whether John Snow's removal of the hand-pump in Soho precipitated or occurred after the number of cholera infections was falling . de Vos et al.  explore these issues for the first time among PWIDs—utilising the Amsterdam longitudinal cohort.
Put simply, de Vos et al.  use modelling to show that large decreases in HIV and HCV incidence could have occurred in Amsterdam solely as a result of natural waning of transmission as infection spreads within the population from higher- to lower-risk individuals, and higher-risk infectious individuals are depleted because of illness. Assortative mixing exaggerates these effects owing to lower-risk PWIDs mainly mixing with themselves, and so naturally slowing infection spread. However, although the observed declines in HIV and HCV incidence could be due to natural progression of the epidemic, large reductions in injecting risk behaviour (model assumes sharing frequency reduces by ≥90% amongst PWID between 1995 and 2000), likely due to scale-up of interventions, were required to reproduce the observed decline in HCV prevalence. The findings have two important implications.
Firstly, they remind us of problems of confounding and ecological fallacy in interpreting observational data: that decreases in incidence or prevalence do not necessarily imply intervention impact and that the evidence base for intervention impact amongst PWID needs to be strengthened [9-12]. This is crucial for understanding previous studies that have observed ecological incidence trends and suggested that they are evidence for intervention impact [13-17]. Indeed, this analysis suggests greater consideration should be given to alternative explanations for decreases in HIV and HCV incidence observed in longitudinal studies , and care should be taken in attributing the changes to intervention effects. For example, could large decreases in PWID HIV incidence reported from Vancouver  be due to the prevention effect of anti-retroviral treatment (ART), as suggested by Wood et al. , or could they be largely due to demographic changes in the population?—something that was not explored in the original analysis. Indeed, the Vancouver cohort also observed similar large decreases in HCV incidence , something that should not be expected if ART was the main driver for the decline in HIV incidence, possibly suggesting that similar processes may be resulting in the epidemiological trends in Vancouver as in Amsterdam. It is therefore important that similar detailed modelling be applied to Vancouver, to assess whether the assertions claimed by Wood et al.  are the most likely reason for the observed trends in HIV incidence.
Secondly, the study is a reminder of the adage that ‘all models are wrong but some are useful’. This analysis emphasises the value in using incidence and prevalence data on two infections to improve the power that modelling can have in unpicking observed disease trends—without the HCV prevalence trends no reductions in risk behaviour would have been needed to fit the model to available data. But it is important to remember that all modelling has limitations that are a product of the model used and assumptions made. Care should therefore be taken in concluding that all declines in HIV and HCV incidence observed in longitudinal cohorts are largely due to natural progression of the epidemic. Specifically, the model used by de Vos et al. assumed strong assortative mixing (more than 70% of partnerships were assumed to be with PWIDs of the same risk level), a model parameter they had no data on. Limited data on PWID mixing from UK respondent-driven sampling studies suggest that much lower levels of assortative mixing may occur . The model also assumed quite high levels of heterogeneity (10 times difference in sharing frequency between high- and low-risk PWID), another parameter they had little data on and so was fit during model calibration. Importantly, both these parameters could have a large effect on their results because less heterogeneity or assortative mixing will likely result in smaller reductions in HIV/HCV incidence as an epidemic progresses. This is due to lower-risk PWIDs having more comparable risk to higher-risk PWIDs, and a higher chance of mixing with them. Information on injecting duration and transitions between high- and low-risk groups was also uncertain, but could have a strong influence on transmission risk. Further, as the authors suggest, the Amsterdam cohort study was started after opiate substitution therapy and needle and syringe programmes were in place—so any change in injecting risk behaviour and the proportion of injectors that are ‘high’ or ‘low’ risk could have already occurred. Separating out what might happen in the absence of any intervention or what would happen naturally is complicated and subject to many assumptions for which we often lack reliable data or empirical evidence. Nonetheless, the authors have done the field great service in reminding us that we need to improve the science and evidence base for the degree to which interventions for PWID reduce HIV and HCV infection risk.
Declaration of interest