Letter by Hartwig et al Regarding Article, “Evaluation of the Pleiotropic Effects of Statins: A Reanalysis of the Randomized Trial Evidence Using Egger Regression”
To the Editor:
It was with interest that we read the publication by Labos et al1 on applying an adaptation of Egger regression to evaluate pleiotropic effects of drugs. They interpret their reanalysis of randomized controlled trials (RCTs) as indicating that the effect of statins on cardiovascular disease (CVD) and diabetes mellitus are mediated primarily, if not entirely, because of reducing low-density lipoprotein cholesterol (LDL-C) levels. We think that this approach has potential, but here, we point out 2 main conceptual limitations of it, which we think question the authors’ conclusion.
Labos et al essentially did a meta-regression of the randomization status (RS)-outcome association (ie, intention-to-treat analyses) on RS-LDL-C association. By RS, we mean whether participants were assigned to statin or control groups, irrespective of compliance with these assignments. Given that perfect compliance is unlikely in real RCTs, RS and statin intake should not be treated as interchangeable. For example, an intention-to-treat analysis would be expected to be robust against confounding by prerandomization factors, but the same is not true for the association between actual statin intake and the outcome.
Suppose that LDL-C has no causal effect on CVD but that statins nevertheless have a causal effect on CVD (that is not mediated by LDL-C) as depicted in the graphical abstract of Labos et al. In this situation, statin intake confounds the association between LDL-C levels and CVD, and therefore, both the association of RS with LDL-C and the association of the RS with CVD depend on the association of RS with statin intake. This would constitute a violation of the InSIDE assumption (Instrument Strength Independent of Direct Effect), which is a critical assumption of the proposed adaptation of Egger regression.1,2 In this example, InSIDE requires that any effect of the RS on CVD risk that is not mediated by LDL-C levels is independent of the effect of the RS on LDL-C levels.
The second likely limitation is that pleiotropic effects of drugs that belong to the same class may be mediated by common mechanisms. For example, statins competitively inhibit the 3-hydroxy-3-methylglutaryl-coenzyme A reductase enzyme, which leads to reduced synthesis of cholesterol in the liver.3,4 Therefore, if there are pleiotropic effects downstream to 3-hydroxy-3-methylglutaryl-coenzyme A reductase inhibition, such effects would be expected to be common to all preparations of statins. For example, farnesyl diphosphate, which is downstream to 3-hydroxy-3-methylglutaryl-coenzyme A reductase inhibition, is a precursor of cholesterol and involved in the prenylation of specific proteins.5 Given that both on-target (eg, mediated by inhibiting cholesterol synthesis) and pleiotropic effects (eg, mediated by inhibiting formation of prenylated proteins) would depend on a common mechanism (eg, 3-hydroxy-3-methylglutaryl-coenzyme A reductase inhibition by statins), they would be expected to be correlated, thus violating the InSIDE assumption. In other words, although different preparations of statins may have specific pleiotropic effects, we feel it is highly plausible that the InSIDE assumption is violated in this example because RCTs of higher dose statins will not only result in a greater reduction in LDL-C but concurrently a greater potential pleiotropic effect.
We think that Egger regression may be useful to assess pleiotropy of RS (eg, effects of RS on CVD not mediated by statin intake), for example, because of differential care according to RS because of lack of concealment, as long as such effect is not correlated with the RS–statin association (ie, InSIDE holds). However, regarding assessing pleiotropy of the drug itself, the above limitations highlight some important methodological challenges. The adaptation of Egger regression was motivated by the Mendelian randomization (MR)-Egger regression method, which uses summary genetic association results to estimate the causal effect of an exposure on an outcome in an instrumental variable framework, and is robust to bias because of genetic pleiotropy as long as the InSIDE assumption holds.2 Therefore, it is likely that other summary data MR methods that may be applied in the RCT context in order to complement Egger regression. For example, the multivariable MR method has been proposed to deal with bias because of genetic pleiotropy on measured variables.6 It may be possible to use an adapted version of multivariable MR (ie, multivariable Egger regression) to adjust for (or at least mitigate) bias because of nonequivalence between RS and statin intake by including compliance measures as additional moderators in the meta-regression.
Multivariable Egger regression may also be used to deal with pleiotropic effects that have a common mechanism, as long as the mediators of such pleiotropic effects were accurately measured in the original RCTs and data on the association between RS and such mediators are available. In practice, however, it seems unlikely that such data would be generally available, given that the primary goal of RCTs is to assess the effect of the drug itself on a given disease outcome, rather than the mechanisms that bring about such effects. In this case, the knowledge of the biological pathways influenced by the drug would be key to assess how likely it is that InSIDE holds.
It seems more likely that InSIDE holds when considering different classes of drugs that influence a common intermediate through different mechanisms. For example, a large meta-analysis of RCTs of antihypertensive drugs on coronary heart disease and stroke showed that, in general, different classes of antihypertensive drugs (and thus subjected to different sources of pleiotropy7) were similarly effective in reducing the risk of CVD, with effects proportional to the reduction in blood pressure.8 Although that study did not use the proposed Egger regression adaptation and may, thus, be prone to bias because of pleiotropic effects,7 in our opinion, it provides more compelling evidence of antihypertensives reducing CVD solely via their effect on blood pressure than the analysis of Labos et al regarding the effect of statins on CVD (or diabetes mellitus) being solely via LDL-C.
We are confident that the proposed adaptation of Egger regression, as well as adaptations of other summary data MR methods, will be important aids in assessing pleiotropic effects of drugs. However, we think that there are still important challenges that require further methodological development before this approach can be fruitfully applied.
Fernando Pires Hartwig
Postgraduate Programme in Epidemiology
Federal University of Pelotas
Maria Carolina Borges
Medical Research Council Integrative Epidemiology Unit
University of Bristol
Debbie A. Lawlor
Medical Research Council Integrative Epidemiology Unit
University of Bristol
Sources of Funding
The Medical Research Council (MRC) and the University of Bristol fund the MRC Integrative Epidemiology Unit (MC_UU_12013/1, MC_UU_12013/5) and the National Institute for Health Research Biomedical Research Centre at University Hospitals Bristol National Health Service Foundation Trust and the University of Bristol. M.C. Borges is supported by the MRC Skills Development Fellowship (MR/P014054/1).
D.A. Lawlor has received support from government and charity funders, Roche Diagnostics, and Medtronic for research unrelated to this correspondence. The other authors report no conflicts.
Letters to the Editor will be published, if suitable, as space permits. They should not exceed 1000 words (typed double-spaced) in length and may be subject to editing or abridgment.
The views expressed in this publication are those of the authors and not necessarily those of the MRC, NHS, NIHR, or the Department of Health.
- © 2018 American Heart Association, Inc.
- Labos C,
- Brophy JM,
- Smith GD,
- Sniderman AD,
- Thanassoulis G
- Lawlor DA,
- Tilling K,
- Davey Smith G
- Law MR,
- Morris JK,
- Wald NJ