Answers are grounded in this paper and cite their sources.
Randomized controlled trials (RCTs) sit at the pinnacle of the hierarchy of evidence, but conducting an RCT is not always practical or ethical. We must be cautious about using observational studies as a basis for action, since the complex network of causes in the real world means that it is impossible to entirely eliminate confounding factors. Correlation, as they say, is not causation. But what if there were a powerful method—drawing strength from both RCT and observational concepts—that was less influenced by confounding? As an example: Is alcohol really good for your heart, if taken in moderation? Some observational studies have suggested so [1]. Before telling patients to add a bottle of wine to their weekly shop, clinicians are rightly cautious. The first problem is confounding. Many confounders cannot be measured accurately (e.g. smoking exposure) or at all (attitudes to health). The second problem is reverse causation: people may cut down on alcohol because of their illness. These problems can make observational studies difficult to interpret. Imagine we were to assign thousands of newborns randomly to high or low alcohol intake over their whole lives. Then we could be more certain that different rates of heart disease are directly due to the difference in alcohol intake (causal effect). This conclusion would require three conditions to be true. First, the imaginary intervention would have to ensure that individuals really did have high or low alcohol intake over their lives. Second, assignment must be truly random as in a well-conducted RCT, thereby eliminating confounding. Third, the intervention must specifically affect alcohol intake, but not heart disease directly or its other risk factors. If it accidentally reduced smoking rates, then we would not know whether effects were due to alcohol or smoking. This thought experiment illustrates the three assumptions (relevance, independence, exclusion) behind the instrumental variable approach to studying causal effects [2], of which Mendelian randomization (MR) is one type. MR, like an RCT, draws its strength from random allocation [3] (Fig. 1). At conception, we were each randomly allocated genetic variants that not only influence physiological factors like urate and cholesterol levels, but also behaviours like tendency to drink alcohol or smoke. MR allows us to infer causation with less concern for confounding or—since genes are not affected by what we chose to do in life—reverse causation. When this was done for the alcohol and heart disease question, for example, it was found that genetically predicted alcohol intake is not protective [1]. MR also contradicted observational associations and forecasted null effects for other risk factors (e.g. high-density lipoprotein (HDL) cholesterol, vitamin D or E for primary disease prevention) subsequently confirmed by very expensive RCTs. Mendelian randomization as an analogue of randomized controlled trials Trial participants are randomly assigned to intervention or control and followed up. In an ideal trial, differences in outcome at the end of follow-up is the causal effect of intervention. According to Mendel’s Law, parental genes are randomly and independently (i.e. inheritance of one trait is independent of another) assigned to egg/sperm. Analogous to an RCT, we were each randomly allocated variants that affect our life-long exposure to risk factors. Under certain assumptions, difference in outcome is the causal effect of genetically predicted exposure levels over the life course. In both RCT and MR, confounders are balanced between groups. Mendelian randomization as an analogue of randomized controlled trials Trial participants are randomly assigned to intervention or control and followed up. In an ideal trial, differences in outcome at the end of follow-up is the causal effect of intervention. According to Mendel’s Law, parental genes are randomly and independently (i.e. inheritance of one trait is independent of another) assigned to egg/sperm. Analogous to an RCT, we were each randomly allocated variants that affect our life-long exposure to risk factors. Under certain assumptions, difference in outcome is the causal effect of genetically predicted exposure levels over the life course. In both RCT and MR, confounders are balanced between groups. Traditional MR studies required genetic information, exposure status and disease outcome to be measured in one large population; this is known as one-sample MR. The relationship between genetically predicted exposure and disease outcome is causal under the above assumptions. However, finding a suitably large population with which to generate these data is no easy task. A more efficient MR approach is to use two separate cohorts sampled from the same underlying population [4]: one for genetic variants of the exposure and another for variants of the outcome, i.e. two-sample MR. Combining their data is possible because genome-wide association study (GWAS) technologies are generally similar across different studies. Summary data (i.e. coefficients summarizing associations between genetic variants and phenotype) from large GWAS are openly available for thousands of phenotypes. For example, does genetic predisposition to putting on weight cause gout [5]? Find a GWAS for adiposity traits and another for gout. Does genetically predicted adiponectin cause lupus [6]? Mix and match, ad infinitum. There are now online tools [7] that allow anyone to run naïve analyses. The number of publications is set to explode; this is one reason why clinicians will need to understand how to interpret and critically appraise MR studies [2]. Another reason is the wide range of clinically important questions that MR can help address. MR is often applied to epidemiological questions, but it is also relevant for aetiological and translational research. MR can help identify and prioritize potential therapeutic targets for drug development [8], find drug repurposing opportunities, and predict adverse drug events [9, 10]. For example, romosozumab is effective for osteoporosis management, but trials suggested an increase in cardiovascular risk (adverse events notoriously challenging to study in RCTs due to substantial person-years of follow-up required). Concern for cardiovascular risk was supported by an MR study of sclerostin that showed increased risk of major adverse cardiovascular events, and also diabetes, hypertension and adiposity [11]. Similar cardiovascular risk concerns were initially raised for IL-6 inhibition due to its association with dyslipidaemia [12]. MR of IL-6 signalling provided reassurance that concurred with RCT [13]. The greater statistical power of the MR study in fact suggested reduced cardiovascular risk with lower IL-6 levels, but instead warned of novel safety signals for atopic diseases like asthma [12]. Supplementary Table S1 (available at Rheumatology online) provides some other recent examples in rheumatology. MR studies have their limitations. Results can be misleading if the core assumptions are violated. However, not all assumptions are empirically verifiable, so sensitivity analyses to support their plausibility are essential [14]. Even if assumptions can be guaranteed, interpretation can still be nuanced. Genetically predicted variation in the exposure can be small; therefore, null results do not necessarily exclude causal effects of pathological levels. On the other hand, positive findings from MR studies may have exaggerated effect sizes; MR tests the causal effect of lifelong exposure on the outcome, which is unlikely to be the same as an intervention in middle age. In summary, MR is an invaluable and increasingly used method to examine causal relationships between modifiable risk factors and disease. It is incredibly versatile—useful for traditional epidemiological questions as well as translational research such as drug development. Readers should scrutinize the assumptions on which conclusions are based. Where possible, we should avoid relying on any single study design, even RCTs, to determine clinical practice. Clinicians should familiarize themselves with MR, which provides causal evidence-base at the interface between RCTs and traditional observational epidemiology. We will all need to learn to speak the language of these studies in order to practice evidence-based medicine in the 21st century. Funding: J.Z. is funded by a Vice-Chancellor’s Fellowship from the University of Bristol. This research was also funded by the UK Medical Research Council Integrative Epidemiology Unit (MC_UU_00011/1 and MC_UU_00011/4). Disclosure statement: J.Z. is employed on a grant funded by GlaxoSmithKline. S.L.M.: Consultancy on behalf of institution to: Roche/Chugai, Sanofi. Investigator on industry-sponsored clinical trials for: GSK, Sanofi. Attendance at EULAR2019 supported by Roche. Co-applicant on research grants from Vifor and Roche. Patron of the charity PMRGCAuk. This publication presents independent research supported by the Leeds Biomedical Research Centre. The views expressed are those of the authors and not necessarily those of the NHS, the NIHR or the Department of Health and Social Care. Supplementary data are available at Rheumatology online.
No takes yet. Share an insight, caveat, or question.
Zhao et al. (2021) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: