Abstract
Mendelian randomization refers to an analytic approach to assess the causality of an observed association between a modifiable exposure or risk factor and a clinically relevant outcome. It presents a valuable tool, especially when randomized controlled trials to examine causality are not feasible and observational studies provide biased associations because of confounding or reverse causality. These issues are addressed by using genetic variants as instrumental variables for the tested exposure: the alleles of this exposure–associated genetic variant are randomly allocated and not subject to reverse causation. This, together with the wide availability of published genetic associations to screen for suitable genetic instrumental variables make Mendelian randomization a time- and cost-efficient approach and contribute to its increasing popularity for assessing and screening for potentially causal associations. An observed association between the genetic instrumental variable and the outcome supports the hypothesis that the exposure in question is causally related to the outcome. This review provides an overview of the Mendelian randomization method, addresses assumptions and implications, and includes illustrative examples. We also discuss special issues in nephrology, such as inverse risk factor associations in advanced disease, and outline opportunities to design Mendelian randomization studies around kidney function and disease.
One of the major aims of medical research is to identify exposures, also called risk factors or intermediate phenotypes, which are causal to the manifestation of a specific outcome, such as disease initiation, disease progression, or response to therapy. Once identified, causal risk factors can enable preventive measures and represent attractive therapeutic targets. Randomized, controlled trials (RCTs) are the gold standard to establish causal relationships.1–3 Proper randomization ensures that study groups are comparable in all characteristics, except for the exposure of interest, which often is a therapeutic intervention. Differences in the outcome can then be directly assigned to the effect of the exposure. However, RCTs cannot always be conducted, because they can be excessively costly, impractical, or even unethical.2,3
When RCTs are not feasible or in addition to RCTs, exposures are often investigated in observational studies.1 Here, study groups usually differ in not only the exposure of interest but also in several observed and unobserved characteristics. Differences in the outcome between exposure groups may be attributed to any of these characteristics or a combination thereof, and observational studies can, therefore, not directly establish causality: an observed exposure-outcome association may not reflect a causal relationship but may arise as the result of confounding2–4 or reverse causation.5 Confounding is often addressed statistically by including known and measured confounders into regression models (multivariable regression). However, when confounders are unobserved, because they are unmeasured or unknown, or when the number of confounders is too large, regression methods may fail to provide unbiased estimates of the true association between exposure and outcome.1
The instrumental variable method was proposed as an alternative statistical method to examine causality of exposure-outcome associations while controlling for any confounder. The concept was first introduced by econometricians almost a century ago and later adopted by medical statistics.6,7 An instrumental variable is chosen to mimic the randomized allocation of individuals to the exposure and thus, ensure comparability of groups with respect to any known and unknown confounder. When such a valid instrument is available, the effect of the exposure on the outcome can be unbiasedly estimated, and thus, causality of an observed association can be assessed.8
Although the instrument could be essentially any variable, genetic variants, such as single-nucleotide polymorphisms (SNPs), are being increasingly used, because their alleles are assigned to individuals before any exposure or outcome. In fact, genetic instruments are nonmodifiable, ensuring lifelong exposure and mitigating concerns about reverse causation.3 During human gamete formation, the alleles of a given SNP are randomly allocated to egg/sperm cells. Consequently, inherited variants are independent of potentially confounding environmental exposures.2,3,9 Because of the relation to Mendel’s Laws, the term Mendelian randomization (MR) was coined.2,10 Many candidate gene and genome–wide association studies (GWAS) have been published over the last decade, which now allow for the conduct of MR studies that exploit these reported associations without the need to recruit new patients or design additional studies. This is reflected in the increasing number of instrumental variable analysis in general and MR studies in particular in the medical literature (Figure 1).3,11
Use of MR and instrumental variable approaches in the literature increases over time. PubMed Search strategy (June of 2016): for MR analysis, “mendelian random*[tiab]” or “Mendelian Randomization Analysis” (medical subject headings [MeSH]); for instrumental variable analysis or MR analysis, “instrumental variable*[tiab],” “mendelian random*[tiab],” or “Mendelian Randomization Analysis” (MeSH). Note that the MeSH term “Mendelian Randomization Analysis” was introduced by MEDLINE in 2010.
The aim of this review is to provide an overview of the MR method, its assumptions, and its analysis steps, including illustrative examples of medical relevance with a focus on the field of nephrology. Special considerations when using the method and implications for medical research are also discussed. Table 1 contains a glossary of terms commonly used in MR studies.
Glossary of terms commonly used in MR studies
MR: Choice of Instrument and Core Assumptions
The choice of the genetic instrumental variable (GIV) is essential to a successful MR study. To allow unbiased estimation of the causal effect of the exposure on the outcome, a valid GIV fulfills three core assumptions (Figure 2A).2,12
Conceptual illustration of the MR method and its three underlying core assumptions as directed acyclic graphs. (A) Conceptual model. (B) Assumption 1. (C) Assumption 2. (D) Assumption 3.
(1) It must be reproducibly and strongly associated with the exposure (Figure 2B).
(2) It must not be associated with confounders (i.e., factors that confound the relationship between exposure and outcome) (Figure 2C).
(3) It is only associated with the outcome through the exposure (i.e., it is independent of the outcome given the exposure) (Figure 2D).
A GIV can be identified by scanning published databases or reports evaluating genetic associations with the exposure of interest.9 The many GWASs conducted over the past decade are a useful resource, because they represent hypothesis-free scans, where the exposure and/or the outcome are tested for association with millions of SNPs.
For illustration, we consider the work by Smith et al.,13 which was an MR study conducted to study whether elevated LDL cholesterol (exposure) is causally related to incident aortic stenosis (outcome). For the identification of a GIV, the authors took advantage of published results from a GWAS of LDL cholesterol levels and identified 31 independent SNPs strongly associated with LDL cholesterol levels (core assumption 1 in Figure 2B): they selected SNPs that reached genome-wide significance (P<5×10−8).13
When multiple candidate GIVs are available, such as when many genetic variants associated with the exposure are known, it is preferable to select those that are located in genes with biologic function that is best understood.3,14 A well understood biologic mechanism simplifies the ascertainment of the second and third assumptions of MR. For instance, genetic variants located in the genes encoding the LDL receptor or the 3-hydroxy-3-methyl-glutaryl-CoA (HMG-CoA) reductase enzyme might be reasonable choices because of their well studied role in LDL cholesterol metabolism. An alternative is the use of different GIVs and the comparison of analysis results obtained for each of them15–17 or the generation of a genetic score composed of multiple GIVs as performed by Smith et al.13
MR: Analytic Method
An MR analysis comprises of two main steps: first, the examination of the three underlying core assumptions and second, the evaluation of the causal effect between exposure and outcome.
Table 2 provides an overview of how to assess the three core assumptions. Only the first assumption of GIV-exposure association (Figure 2B) can be directly tested using the data available for the MR study.3,18,19 The second (Figure 2C) and third (Figure 2D) assumptions are essentially not empirically verifiable.3,14,20 However, the second assumption of no association between GIV and confounders is often considered fulfilled because of the random allocation of alleles to gametes.2 To some extent, this assumption can be tested empirically by assessing associations between the GIV and observed confounders when available.3,19 However, the absence of such associations cannot be considered proof that confounding is absent.3 Statistical tests to address specific threats related to the third assumption have also been proposed.21 Table 3 provides an overview of selected threats that can lead to the violation of the three core assumptions. When available, Table 3 also lists precautions that can be taken to address such threats. Each threat should be considered in the context of the specific research field, and some considerations that pertain to the field of nephrology are discussed below.
The three core assumptions underlying MR and their assessment
Selected threats to inference from MR studies and potential solutions
Different methods have been proposed to carry out the actual MR analysis and estimate the magnitude of causal effects,2,20,22–25 with the choice of method depending on the practical setting. Because the presentation of all methods to estimate causal effects is beyond the scope of this article, the general idea is only briefly described for the simple example that exposure and outcome are both continuous, such as C-reactive protein (CRP) and body mass index. Assuming that all associations shown in Figure 2A are linear and that there are no interactions, a standard approach for effect estimation is a linear model. The causal effect of the exposure (X is CRP) on the outcome (Y is body mass index) via the GIV (G) can then be estimated by , where
(known as the Wald ratio estimate) represents the causal effect estimate obtained from
and
, the regression coefficients obtained from the regression of the outcome on the GIV and the regression of the exposure on the GIV, respectively. In this example, this approach is equivalent to the commonly used two–stage least squares approach, where predicted values from the exposure-GIV regression (first stage) are then regressed against the outcome (second stage).23,26 When the outcome is binary, as in the aortic valve disease example above,13 the linear model may still be used as an approximation.23 Otherwise, methods for nonlinear outcomes are available.2,23–25,27–30
A recent systematic review provides a comprehensive overview of applied methods.31 Accordingly, another commonly used approach relies on the comparison of observed and expected effect estimates for the association between the GIV and outcome. While the observed effect estimate () is obtained from the regression of the outcome (Y) on the GIV (G), the expected effect estimate is calculated as the product of the effect estimates obtained from the regression of the exposure (X) on the GIV (
) and the outcome on the exposure (
).12,31,32
The estimation of the magnitude of the causal effect is not always of interest or may not be possible when, for example, regression results are obtained from studies that have used different data transformations, resulting in incompatible effect estimates. Still, MR can be used to evaluate whether a causal relation exists assuming that the three core assumptions are fulfilled.2
Table 4 provides a summary of all steps necessary to carry out a proper MR study. Each step should be addressed in the resulting publication, allowing readers to assess its validity; special reporting guidance for MR articles is available.18,19,31
Steps of an MR study
Extensions of MR
In recent years, -omics technologies, including genomics, transcriptomics, proteomics, metabolomics, and—more recently—epigenomics, have developed rapidly. The application of these technologies in observational studies has generated a very large number of novel exposures/intermediate phenotypes that researchers can use to assess associations with clinical end points. These scans are so–called hypothesis–free approaches, because they do not rely on underlying biologic assumptions and are, therefore, suited to unravel unknown biology. The results of such association studies represent a vast amount of unbiased information on potentially (modifiable) exposures and GIVs, which can subsequently be used to assess novel causal relationships or verify those examined in RCTs. To address this new situation, several extensions to MR approaches have been developed in recent years.3,11,33 These extensions are likely to receive more and more recognition. Here, some of these extensions are briefly described.
MR approaches that use summary statistics facilitate the use of results from published GWAS studies and do not require a separate study in which to carry out the MR analysis. In this approach, ratios on the basis of published summary statistics (regression coefficients that represent the numerator and denominator of the above formula) are used to assess the causal effect of an exposure represented by a single or multiple GIVs.16 Two-sample MR refers to an approach where the association estimates (summary statistics) between the genotype and the exposure and between the genotype and the outcome are generated or collected from two different datasets without or only with a limited number of overlapping individuals.34,35
Bidirectional MR allows for examining an observed exposure-outcome association from both sides. If valid GIVs are available for both exposure and outcome, this approach can help to address the directionality of a causal association,36,37 because reverse causation is a common problem in observational epidemiology.
Network MR can be used to investigate more complex causal relationships between variables, such as when some of an exposure’s effect on the outcome occurs through a mediator variable.38 The simplest network can be investigated by a two-step MR, where a GIV for the exposure is used to estimate the causal effect of exposure on the mediator variable in a first step, and another GIV for the mediator variable is used to estimate the effect of the mediator on the outcome in a second step.3,39,40
Finally, MR will be increasingly used in a hypothesis-generating manner: by testing all pairwise relationships within large multidimensional datasets, associations can be identified that are then followed up to test specific hypotheses about causality in an MR setting.3,41
Use of MR in Cardiovascular Disease: Examples
One area of research in which MR studies have contributed important insights is the potential causality of factors associated with coronary heart disease (CHD) or cardiovascular disease (CVD) more generally. Over the past years, several large, well powered studies arrived at the conclusion that elevated blood concentrations of LDL cholesterol are causally related to the development and progression of CHD and other CVD outcomes.5,13,42–44 Because of the existing biologic knowledge about LDL cholesterol metabolism, genetic variants that act on a specific protein or portion of the metabolic pathway can be compared with the effect of a drug acting on the same target.11 It is, therefore, possible to compare genetic effects on LDL cholesterol concentrations with the effects of RCTs of statins, effective and widely used medications to lower blood LDL cholesterol (Figure 3). The larger CHD risk reduction conferred by GIVs compared with statins for an equal level of LDL cholesterol reduction can be explained by the lifelong exposure to LDL cholesterol–lowering genetic variants.
Risk reduction of CHD associated with LDL cholesterol–lowering genetic variants compared with LDL cholesterol–lowering medications. A steeper risk relationship with the genetic variants than with medications for the same amount of lower cholesterol reflects the effects of lifelong exposure to genotypes. Effect estimates reflecting the risk reduction in CHD for the individual clinical trials were taken from Webfigure 3i of a meta-analysis of randomized trials of statins.85 Effect estimates for the genetic variants, the genetic and clinical summary effects, and 95% confidence intervals were taken from Table 1 and Figure 5 of the work by Ference et al.43
This example also illustrates that unbiased genetic screens, such as GWASs, coupled to MR approaches can be used to inform drug development, which aims at pharmacologic modulation of causal risk factors. Common CHD–associated genetic variants identified in genome–wide screens map into different genes affecting LDL cholesterol metabolism, among them the genes encoding the HMG-CoA reductase,45 the target of statins, and proprotein convertase subtilisin/kexin type 9,46 the target of a new class of LDL cholesterol–lowering drugs.47 These therapies have already been developed but exemplify that GWASs coupled to support for a causal association between LDL cholesterol and CHD from MR studies can be useful to identify additional novel therapeutic targets. This is true even if the common genetic variants identified from GWASs only have a small effect on the exposure, which is the case for common variants in the gene encoding HMG-CoA reductase and LDL cholesterol concentrations.43
MR was also used to evaluate the potential causality of observed associations between other risk markers and CVD. These studies found that the association between lower HDL cholesterol concentrations and myocardial infarction as well as between higher concentrations of plasma CRP and CVD are unlikely to be causal.44,48–50 Therefore, increasing HDL cholesterol concentrations or lowering CRP concentrations are unlikely to represent effective therapeutic approaches to reduce CVD risk. Because of the difficulty in verifying the MR core assumptions, some authors have, in fact, suggested that negative rather than positive MR studies may provide the most reliable information and could, consequently, be used to identify targets on which additional drug development is unlikely to be successful.51
Another application of MR is to inform drug development with respect to side effects. MR can be used to obtain information about causality of observed side effects as exemplified in the case of statins and risk of diabetes.52 Several lines of evidence from observational studies and clinical trials of statins have shown an increased risk of type 2 diabetes in individuals receiving statin treatment. MR was then used to show that a GIV in the gene encoding HMG-CoA reductase was associated with increased risk of diabetes.52 This observation suggests that increased diabetes risk in statin users may be an on-target effect directly mediated by reduced HMG-CoA reductase activity. Therefore, the attempts to develop improved drugs that more specifically target HMG-CoA reductase to reduce off–target side effects are unlikely to be effective with respect to diabetes risk.11
Use of MR in Nephrology: Examples
In nephrology, MR studies have been used to study the causality of the association between serum urate concentrations and CKD. One relatively small study presented evidence for a causal role of serum urate in CKD progression by using the SNP rs734553 in the urate transporter gene SLC2A9 as a GIV.53 This SNP had previously shown very strong association with uric acid levels, even in small population samples.53 However, the GIV was not associated with the exposure, serum urate, in this CKD sample. This example illustrates that, in patients with CKD, a well characterized association, such as that between a GIV in SLC2A9 and serum urate, may be masked by the effect of low eGFR on urate concentrations and the high intake of urate-lowering medications. In contrast to this first study, another MR study addressing a similar question reported support for a causal relationship between higher serum urate concentrations and higher eGFR, i.e., better kidney function.54 In addition, several studies found no association between urate-associated GIVs and eGFR or AKI.55,56 The potential causality of the observed association between serum urate concentrations and adverse renal outcomes, therefore, deserves additional study.
In a very recent example, the causal relationship between serum iron levels and eGFR in the general population was evaluated.57 Although iron depletion is a common consequence of CKD, it was unclear whether iron itself could affect kidney function. Strong GIVs were selected from GWASs of iron and ferritin levels58 and then, tested for association with eGFR using GWAS summary results on the basis of 74,000 individuals in the CKD Genetics (CKDGen) Consortium.59 The authors showed that, in the general population, lower iron and ferritin levels were associated with a statistically significant albeit small effect on lower kidney function.59 Whether such a small effect is relevant to the long-term maintenance of renal function requires additional study.
Support for causal associations with progression of type 1 diabetic nephropathy has been described for both the protein KIM-1 in urine and obesity.60,61 The latter study represents an illustrative example of a comprehensive MR study with nephrologic outcomes. No support for a causal association was reported from an MR study investigating blood concentrations of fetuin-A and mortality in patients on dialysis.62
Previous review articles have provided an overview of additional examples of MR and specific issues arising in the field of nephrology.12,63 In particular, inverse risk factor associations and survival bias should receive special attention. Inverse risk factor associations describe the phenomenon that associations observed in the general population or patients with early CKD are reversed in direction in patients on dialysis. For example, higher cholesterol concentrations are associated with lower risk of mortality in patients on dialysis, which may be attributed to the cholesterol-lowering effect of systemic inflammation and malnutrition.64 Such strong influences of the environmental context need to be taken into account when planning MR studies in the setting of ESRD or any other disease–based study population. An instrument that is valid in a population-based setting might not be valid in ascertained populations when it influences factors that are associated with the choice of ascertainment. Consequently, in ascertained populations, the same core assumptions for a valid GIV must apply and should be verified with same care. Another issue in MR studies in nephrology is that many patients with CKD die from CVD causes before reaching ESRD. If these CVD end points are associated with the GIVs studied in an ESRD population, survival bias may influence the results.65
Much of the work to identify genetic underpinnings of complex diseases is undertaken by large international collaborations, such as the CKDGen59,66–68 and the Asian Genetic Epidemiology Network69 Consortia for kidney function in health and disease, efforts to study IgA nephropathy,70–72 the Global Urate Genetics Consortium (GUGC) Consortium for serum urate concentrations and gout,73 and several efforts to study diabetic nephropathy in both types 174 and 2 diabetes.75 The full genome–wide summary results for some of these studies are publicly available (Table 5). These datasets represent a valuable resource for the conduct of MR studies in nephrology: they allow for both identifying GIVs for kidney function to be tested for association with other outcomes and evaluation of GIVs that represent other kidney function–related traits on measures of kidney function and disease.
Helpful resources for design and conduct of MR studies
The CKDGen and the GUGC Consortia have already conducted projects that include elements of MR to address the relation of kidney function with CVD76 and the relation of serum urate concentrations with hypertension and components of the metabolic syndrome.55,73 A complicating factor in the past, the unavailability of sufficiently strong GIVs, is now changing with the emergence of consortium data from large GWAS meta–analyses that allow for the construction of genetic risk scores.68 MR techniques can, therefore, now be used to study one of the most common and complex problems in nephrology, the relationship between reduced kidney function and cardiovascular risk factors, morbidity, and mortality.77 The publicly available data and tools summarized in Table 5 provide a basis to initiate such studies.
Considerations and Future Directions
It is interesting to speculate to what degree MR will be used in the future. In contrast to observational studies that often suffer from biased results because of (unmeasured) confounding and in contrast to some other analytical methods, MR can deal with any confounding by design as long as a valid GIV is available.78–81 However, a valid instrument may not be available for every research question because of lack of knowledge, and publicly available data sources may not always provide information on the associations of interest.
Another potential limitation in MR studies is the statistical power of the study design. The power of an MR study depends on several aspects, including the proportion of variance in the exposure explained by the GIV and the magnitude of the causal association between exposure and outcome. Formulas to perform sample size calculations during planning of an MR study are available.82–84
Despite these limitations, we believe that MR approaches will increasingly be used to assess causality of risk factor associations in medical research. Other reasons that support the use of MR are in the nature of genetic variants being appropriate instruments, because they act as lifelong (fixed) exposures, and there is little concern about confounding and reverse causality. Moreover, modern laboratory techniques allow for measuring genetic variants with very little error compared with other measurements in observational studies.3 Finally, continued improvement of our understanding of pathophysiologic mechanisms and the public availability of summary results from GWASs together with the methodologic extensions of MR will give raise to a multitude of new research questions that might be addressed using MR.
Conclusions
The validity of results from MR studies depends on the correctness of several assumptions, which should be carefully checked and interpreted in the context of prior (biologic) knowledge. If applied correctly, MR can be highly useful to inform drug development and repurposing. Whenever possible, however, causal associations should be confirmed in an RCT. The use of MR as a promising approach to assess the causality of observed exposure-outcome associations through genetic instrumental variables will likely become increasingly popular in medical research in general as well as in nephrology.
Disclosures
None.
Acknowledgments
The work of A.K. was supported by German Research Foundation grant CRC 1140.
Footnotes
Published online ahead of print. Publication date available at www.jasn.org.
- Copyright © 2016 by the American Society of Nephrology