Abstract
Data from the REPRIEVE trial have been used to study the association between treatment with abacavir and cardiovascular disease in people with HIV. Past exposure to abacavir before the trial was found to be just as risky as exposure during the trial. This calls into doubt the widely accepted hypothesis that recent exposure to abacavir increases the risk of cardiovascular disease but that risk rapidly attenuates after exposure ends. The evidence for this hypothesis is weak while evidence for an alternative hypothesis has been neglected. An alternative hypothesis is that risk from exposure to abacavir is not immediate but increases as exposure cumulates, plateauing after about three years. Confusion over the likely effect of abacavir has arisen partly because overly simplistic methods have been used to model the exposure outcome relationship in observational cohort data. Better methods are available and should be used in future to avoid misdirected research effort.
It seems that those who do not learn from past research are doomed to repeat it. People with HIV have a higher risk of cardiovascular disease (CVD) and data from the REPRIEVE trial have been used to study whether treatment with abacavir further increases that risk.1,2 Although such studies are observational, the data themselves were collected in a randomised trial focused on CVD outcomes and in those at low to moderate risk of CVD, and are therefore valuable for exploring the association between abacavir and CVD. In both studies investigators conclude there is an association but found themselves with a conundrum. Why would past exposure before the trial appear to be just as risky as exposure during the trial if the effect of abacavir on CVD is supposed to dissipate rapidly after exposure ends? The issue of how to represent exposure in analyses of observational data has beset research on this topic since the very first paper.
The “reversible exposure” hypothesis is widely accepted. 3 It holds that recent exposure to abacavir – in the preceding 6 months – increases the risk of CVD; that risk does not increase with cumulating exposure and is reversible, such that there is no residual risk 6 months after exposure ends. This hypothesis arose out of a study where three models were fitted in sequence, with exposure represented by one to three variables: linear cumulative exposure, recent exposure and past exposure. 4 While the two indicator variables – recent and past exposure – were likely, though not necessarily correlated with each other, both had to be correlated with cumulative exposure because they were components of it. So when a recent exposure variable was added to a model containing a cumulative exposure variable, the effect of cumulative exposure appeared less important; a classic consequence of collinearity that led investigators to conclude that risk was related to recent exposure and not to cumulative exposure. Further, when past exposure was added to a model with the other two variables – and was not significant – investigators concluded that the risk was reversible.
This was very weak evidence for reversibility given the dependencies between the three variables and that the effect of past exposure to abacavir was imprecisely estimated (relative risk 1.29, 95% confidence interval 0.94-1.77 4 ). It is difficult to untangle the separate effects of dependent exposures. If the estimates for the various effects of exposure are correlated, estimates become unstable and confidence intervals are inflated. 5 In linear models, this only occurs when there is a high correlation between the variables representing exposure. Non-linear models – such as Cox models and logistic and Poisson regression – are more complicated and correlation between estimates cannot be predicted simply by assessing the correlation between the exposure variables themselves. 5 Unstable estimates can result even when the correlation between exposures variables is far more modest, if there is little variation among exposures across a cohort or if few events are seen for some combinations of the exposure variables.6,7 The other problem is that including cumulative exposure linearly in a model equates to an implausible assumption that risk cumulates indefinitely. More plausibly “a threshold amount of exposure may be required for risk, or risk may level off after a certain amount of exposure” and therefore “investigators considering cumulative exposure should allow for nonlinear relationships”. 8
A plausible “cumulating exposure” hypothesis, suggested by alternative analyses, is that the risk of CVD given exposure to abacavir is not immediate but increases as exposure cumulates; not indefinitely, rather the risk plateaus after about three years exposure and then continues at this elevated level until exposure ends. This implies that when exposure ends, risk abates only slowly because past exposure within the last three years still influences current risk. 9 This hypothesis arose out of our study where the exposure outcome relationship was modelled as a flexible function of past exposures. 10 Prompted by our work, whenever investigators looked for evidence of an association between CVD and nonlinear cumulating exposure to abacavir, they found it.11,12 Yet in our study and another, simplistic exposure modelling was shown to lead to the same misleading results as before: cumulative exposure was not significant in a model that included both recent and linear cumulative exposure.10,12
Most investigators didn’t look for evidence of a nonlinear cumulative relationship,13–15 even when it was apparent in their raw data. 16 They fitted models with recent exposure only, not even checking for evidence that the risk was reversible. Clinical evidence for a rapidly acting mechanism or for reversibility has proved elusive. 17 Biomarker studies have not found consistent differences between abacavir and tenofovir based antiretroviral therapy in the levels of biomarkers representing various pathways implicated in CVD pathogenesis.18–21 In one trial, the reversibility of one biomarker among nine does not inspire confidence.22–24 On the other hand, traditional lipid risk factors – total cholesterol, low-density lipoproteins and triglycerides – have consistently been higher when individuals are exposed to abacavir rather than tenofovir.19–22,25,26 Tenofovir may depress lipid levels more than abacavir elevates them; however in a recent study, those switching to dolutegravir and lamivudine had lower cholesterol after 72 weeks than those remaining on dolutegravir, lamivudine and abacavir. 27 Cumulative exposure to abacavir has been associated with subclinical noncalcified coronary plaque.28,29 Abacavir may lead to chronic inflammation and atherosclerotic progression rather than immediately trigger the rupture of vulnerable plaque. 30
Why is this still relevant, since clinicians seldom prescribe abacavir now and certainly not to aging patients with risk factors for CVD 31 ? It matters because under the first hypothesis, when someone stops taking abacavir, their elevated risk rapidly attenuates and ongoing monitoring is unnecessary. Under the second hypothesis, it would be prudent to monitor and to consider a patient’s history of exposure to abacavir in the previous three years when deciding whether to prescribe a statin. 32
The investigators of the REPRIEVE studies are to be congratulated for checking whether past exposure to abacavir was risky. 2 Unfortunately, their past and current exposure variables are not independent: current exposure at study entry was with or without a history of past exposure. This makes it less certain that the two effects have been reliably estimated and so it would be unwise to say more than both estimates imply an association. Note that median past exposure to abacavir was 3 years: half those with past exposure to abacavir had been exposed for three years or more and were, under the “cumulating exposure” hypothesis, potentially at high risk. In the REPRIEVE studies, investigators never considered whether the association was related to non-linear cumulating exposure to abacavir, nor whether the association differed depending on whether trial participants received a statin or placebo. However, the modest number of CVD events in the trial data makes both questions difficult to answer.
Taking a wider view, research on abacavir illustrates how not to use observational cohort data to detect and study associations between drug exposures and adverse events. Perhaps the association between abacavir and CVD was so unexpected that investigators focused on the possibility of unmeasured or time dependent confounding, forgetting that for causal inference the relationship between exposure and outcome must also be correctly specified. 33 Three points are worth emphasising. First, when screening drug exposures for possible associations with an adverse event, use hierarchical multiple exposure modelling because this reduces the chance of false positives.34,35 Second, when focusing on a specific drug or drug class in explanatory modelling, re-parameterise exposure variables so they are not components of each other, to improve the interpretability of parameter estimates. 36 Third, when modelling cumulative exposure, expect nonlinear cumulative relationships, particularly if there has been no signal in drug registration trials (because most drug trial participants will have been exposed for a year or less). An R package is now available for the method we used37,38; this method has the advantage that nonlinear relationships can be modelled without needing to know much in advance about what the relationship might look like. It’s about time – not just to stop using abacavir, 31 but to start using better statistical methods to detect and study associations between drug exposures and adverse events in observational cohort data.
