Introduction to Mendelian randomization

ABSTRACT

Mendelian randomization (MR), i.e. instrumental variable analysis using genetic instruments, is an approach that incorporates population genetics to improve causal inference. Given that genetics are randomly allocated at conception, this resembles the randomization process in randomized controlled trials and hence is more resistant to unobserved confounding compared to conventional observational studies (e.g. cohort studies). The seminar paper briefly described the origin of MR and its underlying assumptions (relevance, independence, and exclusion restriction). This was followed by introducing one sample MR designs (in which instrument-exposure and instrument-outcome associations are derived from the same sample) and one sample MR design (in which instrument-exposure and instrument-outcome associations are derived from different samples). The seminar paper then summarized key aspects of MR studies, such as instrument selection, data sources for conducting MR studies, and statistical analyses. Variations of MR design were also introduced, such as how this design can inform the effect of drug targets (drug target MR). The STROBE-MR checklist and relevant MR guidelines were introduced. The seminar paper concluded by discussing the credibility crisis of MR studies.

INTRODUCTION

Identifying causes of disease is a fundamental objective in epidemiologic research. Whilst randomized controlled trials (RCTs) are considered the gold standard in causal inference due to the use of randomization to reduce confounding, RCTs cannot address all relevant questions, such as exposures that are difficult to modify (e.g. puberty) or exposures where adherence would be challenging (e.g. long-term dietary modification). Hence, observational studies remain a major approach to study disease etiology. However, observational studies are vulnerable to confounding, which could have explained discrepancies between observational studies and RCTs. For example, the inverse association of vitamins with cardiovascular disease in observational studies could be confounded by sociodemographic and lifestyle factors which are difficult to control for¹⁾. Other examples include high-density lipoprotein cholesterol²⁾ and hormone replacement therapy with cardiovascular disease³⁾, where RCTs failed to replicate the “protective” associations in observational studies^4–6). Although there have been several methodological developments to detect and address confounding, including the use of directed acyclic graphs⁷⁾, target trial emulation⁸⁾, E-values⁹⁾, and control outcomes¹⁰⁾, residual confounding remains a key issue given the lack of randomization in observational studies. Similarly, reverse causation would be an issue where preclinical phases of diseases influence both exposures and outcomes, such as reduced body mass index and lipids in preclinical phases of cancer¹¹⁾. Hence, the use of alternative designs which do not rely on no unmeasured confounding assumption and are more resistant to reverse causation would be helpful, where consistent findings across different designs requiring different assumptions can increase the certainty of the evidence of a genuine causal relation (i.e., triangulation of evidence)¹²⁾.

INSTRUMENTAL VARIABLE ANALYSES

Instrumental variable (IV) analysis is a study design which can estimate the causal effect of exposure on outcome even in the presence of confounding¹³⁾. The three underlying assumptions of IV analysis are described below (Fig. 1):

Fig. 1 . Directed acyclic graph depicting the instrumental variable study design.

1. An instrument (Z), also known as an instrumental variable (IV), is associated with the exposure of interest (X) [Relevance];

2. The instrument (Z) and outcome (Y) association is unconfounded [Independence]; and

3. Any association from instrument (Z) to outcome (Y) is via exposure (X) [Exclusion restriction]

If these three assumptions are fulfilled, the IV analysis can test for the presence of a causal effect although the interpretation of the estimate will also depend on the 4^th assumption (homogeneity or monotonicity)¹⁴⁾. Previous IV analyses considered various variables as IVs, such as physician prescribing preferences (to proxy drug treatment) and distance to hospitals (to proxy timing to treatment). However, a main challenge is whether these IVs are valid since any violation of these assumptions would yield incorrect inference, as commented elsewhere¹⁵⁾.

INSTRUMENTAL VARIABLE ANALYSES USING GENETIC INSTRUMENTS (MENDELIAN RANDOMIZATION)

Starting from early 2000, there was a growing interest in using genetics as IVs to improve causal inference in an observational design, i.e. instrumental variable analyses using genetic instruments¹⁶⁾. A major difference between conventional IV and genetic IV is that genetics are randomly allocated at conception based on Mendel’s law, and hence genetic IVs are largely independent of factors which commonly confound the exposure-outcome associations¹⁷⁾. This makes the independence assumption more likely to hold. Given this distinct characteristic where the genetic IV resembles randomization in RCTs, people started to call instrumental variable analysis using genetic instruments as Mendelian randomization (MR)¹⁸⁾.

MR was not commonly implemented initially due to feasibility issues. These included the lack of genetic data in cohort studies, difficulty in identifying strong genetic IVs due to lack of large-scale genome-wide association studies (GWAS), and large same sizes needed to achieve adequate statistical power¹⁹⁾. Hence, earlier MR studies relied on a single cohort and a single instrument chosen based on functional variants strongly related to the exposure (e.g. ALDH2 for alcohol use, FTO for obesity, or CRP for C-reactive protein)^20–22).

In 2015, a seminal paper highlighted the plausibility of using only summary statistics from GWAS to perform MR analyses, which resolved issues in statistical power and resources²³⁾. Alongside with the emergence of large-scale GWAS (and hence stronger instruments)²⁴⁾, the establishment of mega Biobanks^25,26), statistical packages²⁷⁾, and curated GWAS depository²⁸⁾, the number of MR studies has increased exponentially²⁹⁾.

ASSUMPTION OF MENDELIAN RANDOMIZATION

Similar to IV analyses, MR studies rely on the same three main assumptions, i.e. relevance, independence, and exclusion restriction. Below are some scenarios where these assumptions may be violated.

1. Relevance

As instruments are identified from large GWAS using a very stringent p value (p value < 5 × 10⁻⁸), relevance assumption is likely satisfied (For details, please refer to “Consideration for instrument selection” section). However, possible violations of this assumption include a mismatch in ethnic groups between the GWAS used to derive the instrument and the dataset used for MR analyses (e.g. instrument predicts exposure only in one ethnic group but not another), and the application of the instruments in datasets where corresponding exposure status is invariant (e.g. using COVID-19 instruments in data sources where no COVID-19 cases are present).

2. Independence

Genetics are largely unconfounded by conventional confounders and hence the independence assumption is likely valid. However, possible violations of this assumption include confounding by population stratification (ethnic differences in genetic allele frequencies and phenotypic distribution) although this can be controlled by adjusting for principal components³⁰⁾, dynastic effects (confounding by phenotypes that are transmitted from previous generation)³¹⁾, and assortative mating on traits (spurious genetic correlation in traits between mother and father leading to apparent genetic correlation in traits in offspring)³¹⁾. The use of within family MR design can help circumvent these issues³¹⁾.

3. Exclusion Restriction

This is a main threat to MR studies. A main violation of this assumption is including variants which are pleiotropic. Pleiotropic effects can be categorized as vertical and horizontal pleiotropy. Vertical pleiotropy refers to genetic effects on the outcome that are mediated via exposure of interest (e.g. variants predicting smoking also relates to lung cancer risk), and hence would not violate this assumption³²⁾. On the contrary, horizontal pleiotropy refers to the genetic effects on the outcome that are not via exposure of interest (e.g. using variants in gene regions (e.g. HFE)³³⁾ which relate to numerous phenotypes which are unlikely to be solely attributed by the exposure (e.g. iron) and hence can bias subsequent analyses³⁴⁾)³²⁾. The other possibility of violation is selection bias, which can arise from participation bias (e.g. low response rate in UK Biobank)³⁵⁾, and survivor bias (e.g. for disease with late onset)³⁶⁾, collider bias arising from covariable adjustments in corresponding GWAS³⁷⁾, measurement error of exposure³⁸⁾, and time-varying exposures³⁸⁾. Current approach to tackle violation of exclusion restriction is mainly employing statistical analyses which replaced this assumption with other assumptions, such as instrument strength independent of direct effect (InSIDE) assumption for MR-Egger³⁹⁾.

ONE SAMPLE MENDELIAN RANDOMIZATION

One sample MR refers to the MR design where the genetic associations with exposures and outcomes are obtained from the same dataset. Researchers implementing this design need to have access to a single dataset with instruments, exposures, and outcomes measured. For example, in a previous MR study exploring the relation of alcohol use and cognitive function, the authors obtained the genetic variant for alcohol use (rs671), alcohol use and cognitive function in the Guangzhou Biobank Cohort Study (GBCS) and assessed the impact of alcohol use in cognitive function using two-stage least squares approach²⁰⁾. If instruments weakly associated with exposure are used (i.e. weak instrument bias, often defined as F statistics <10), the MR estimate will be biased towards the confounded observational association⁴⁰⁾. Considering statistical power, one sample MR is conducted in large cohort studies, such as HUNT study⁴¹⁾, UK Biobank²⁵⁾, China Kadoorie Biobank²⁶⁾, and AllofUs Research⁴²⁾. A main advantage of individual level one sample MR is the ability to assess non-linearity although there are concerns over the methods⁴³⁾.

TWO SAMPLE MENDELIAN RANDOMIZATION

Two sample MR refers to the MR design where the genetic associations with exposures and outcomes are obtained from different datasets. In brief, researchers only need to obtain summary statistics of the genetic associations with exposures and outcomes in different datasets to conduct this design. For example, in a previous MR study exploring the relation of lung function with cardiometabolic disease risk, the authors obtained the genetic instruments for lung function from genetic summary statistics of the UK Biobank and obtained the genetic associations with cardiometabolic diseases from summary statistics of genetic consortia, and used two sample MR analytical approaches (e.g., inverse variance weighted method) to assess the impact of lung function in cardiometabolic disease risk⁴⁴⁾. In contrary to the impact of weak instrument bias in one sample MR, the MR estimate will be biased towards the null in a two sample MR if the exposure and outcome datasets are completely non-overlapping⁴⁰⁾. However, with increasing overlap, weak instrument bias will increasingly bias the estimate towards the confounded observational association^40,45), although such biases could be potentially corrected⁴⁶⁾. When the exposure and outcome datasets are derived from the exact same study, this would be considered as one sample MR. As two sample MR can be implemented as long as the researchers have access to publicly available summary statistics from large GWAS or relevant Biobanks, which circumvent limitations in resources and statistical power²³⁾, this is now the main type of MR in the literature.

CONSIDERATION FOR INSTRUMENT SELECTION

To conduct either one sample or two sample MR studies, genetic instruments for exposure are necessary. These genetic instruments are normally identified based on large scale GWAS of respective exposure, where single nucleotide polymorphisms (SNPs) across the entire genome strongly correlated with the exposure are reported. For example, if researchers are interested in exploring the health effects of body mass index (BMI), GWAS of BMI should be identified to extract relevant SNPs as genetic instruments⁴⁷⁾. SNPs from GWAS are often selected based on its strength of association with exposure, commonly based on GWAS significant p value threshold (e.g. 5 × 10⁻⁸) although other studies may use a more relaxed cutoff if the number of IVs is limited, owing to smaller exposure GWAS⁴⁸⁾. However, a general rule of thumb is that the instrument should have an F statistic of at least 10 to reduce the risk of instrument bias, which can be approximated using (Beta_ZX/SE_ZX)^{2 49,50)}. Highly correlated SNPs (based on r² (e.g. a cutoff of 0.001) using ethnic specific population reference panel) are also removed to avoid double counting the effect in downstream analyses. In scenarios where GWAS of exposures provided variants with different adjustment models/data sources, the one which is least susceptible to confounding by population stratification (i.e. adjusted for principal components, ethnic specific) and collider bias (model without heritable covariable adjustments) would be preferred³⁷⁾. It is also recommended to extract the instruments from the original GWAS paper (i.e. tables reporting lead SNPs) instead of the corresponding summary statistics, because these statistics may not have included all samples in the original GWAS leading to reduced instruments (e.g. exclusion of 23AndMe data due to terms of agreement between the GWAS collaborators and 23AndMe. See example here: https://conservancy.umn.edu/items/ca7ed549-636b-41c0-ae79-97c57e266417). If the GWAS is from one single study, there is a possibility of winner’s curse (genetic instrument for exposure has inflated estimates)⁵¹⁾. The corresponding MR estimate will be deflated in two sample MR setting but inflated in one sample MR setting although some suggested the impact is not substantial^51,52). If using non-overlapping samples is not plausible, using a split sample approach could be a solution⁵³⁾.

DATA SOURCES FOR CONDUCTING MENDELIAN RANDOMIZATION STUDIES

Genetic associations with the outcomes can be retrieved based on summary statistics from GWAS. These can be often identified based on data depository from genetic consortia (e.g. CARDIoGRAMplusC4D consortium for coronary artery disease, https://www.cardiogramplusc4d.org/)⁵⁴⁾ or the NHGRI-EBI GWAS Catalog (https://www.ebi.ac.uk/gwas/) using accession number reported in the original GWAS⁵⁵⁾. Similar to which datasets the instruments should be selected from, the model/dataset which is least biased and confounded should be prioritized. The data should be of the same ancestry (if possible) as the exposure GWAS, or else this could violate the relevance assumption. When certain genetic instruments identified in the exposure GWAS cannot be identified from the outcome GWAS, proxy genetic variants can be identified either manually via LDlink (https://ldlink.nih.gov/)⁵⁶⁾, or via existing statistical packages (e.g. TwoSampleMR) using the correct ethnicity²⁸⁾. Apart from individually published GWAS, there are also curated databases which report genetic summary statistics across a wide range of phenotypes in large Biobanks, such as the UK Biobank (https://www.nealelab.is/uk-biobank)²⁵⁾, Biobank Japan Project (https://pheweb.jp/)⁵⁷⁾, FinnGen (https://www.finngen.fi/en)⁵⁸⁾, and Integrative Epidemiology Unit (IEU) OpenGWAS project (https://gwas.mrcieu.ac.uk/)⁵⁹⁾. These can be used as either exposures (e.g. blood pressure, lung function) or outcome datasets (e.g. breast cancer). However, caution is needed that some of these databases may not be as updated. For one sample MR analysis using individual level data, this would require access to Biobank data such as UK Biobank or China Kadoorie Biobank which researchers can apply for access (https://www.ukbiobank.ac.uk/, and https://www.ckbiobank.org/).

STATISTICAL ANALYSES

In one sample MR design, two-stage least squares is commonly used, where there are two sets of regressions using the same data source. In brief, researchers will conduct a regression of the exposure on the instruments (first stage), and another regression of the outcome on the fitted values of the exposure obtained in the first stage (second stage) to calculate the estimate, with the use of robust standard errors⁶⁰⁾.

For the more commonly implemented two sample MR design, the main analysis approach is the Wald ratio (for one instrument), or inverse variance weighted method⁶¹⁾. A main assumption for inverse variance weighted method is balanced pleiotropy (horizontal pleiotropic effects across instruments get cancelled out), which is often difficult to verify. Conventionally, researchers will present heterogeneity statistics of estimates from each instrument, either via Cochran’s Q statistics or I², where high heterogeneity indicates presence of invalid instruments, as well as MR-Egger intercept test to examine overall horizontal pleiotropy³⁹⁾. Given it is likely implausible that all instruments are valid, there are alternative two sample MR methods developed which replace the original IV assumptions with other assumptions for valid inference⁶¹⁾. Examples include weighted median⁶²⁾, MR-Egger³⁹⁾, MR-PRESSO⁶³⁾, and weighted mode⁶⁴⁾, where these tests have different assumptions²⁷⁾. Consistencies across sensitivity analyses with different assumptions would add confidence that the observed association is unbiased. In case of discrepancies, more investigations should be followed. For example, did the MR-Egger intercept test indicate possible overall horizontal pleiotropic effect and hence gives different results compared to inverse variance weighted method?³⁹⁾ or are there outliers based on the scatterplot of the instruments which may impact the accuracy of MR-Egger more than other analyses?⁶⁵⁾ In situation where variants predict several correlated phenotypes and hence constitutes significant horizonal pleiotropy (e.g. different lipids)⁶⁶⁾, multivariable MR is often included to reduce biases arising from horizonal pleiotropy. Multivariable MR can also be used in mediation analysis (See section on “Variations of Mendelian randomization design”)⁶⁷⁾. Although these methods have been developed for implementation in the two sample MR setting, a previous review indicated that many of these methods can also be used in one sample MR, apart from MR-Egger unless the I²_ZX is high⁶⁸⁾. A lot of these methods are now being consolidated in R packages such as “TwoSampleMR”²⁸⁾ and “MendelianRandomization”⁶⁹⁾. In particular, TwoSampleMR package links to IEU-OpenGWAS and can streamline data extraction for downstream analyses.

MR studies require larger sample sizes to achieve adequate statistical power given the low variance of the exposure explained by the instruments^19,70). To calculate statistical power, there are online calculators, such as mRnd (https://shiny.cnsgenomics.com/mRnd/)⁷¹⁾, or user written functions⁷²⁾.

Lastly, although two sample MR generally has larger power, a main advantage of one sample MR is that it enables assessment of non-linearity (i.e. non-linear MR). However, the original approach (residual stratification method)⁷³⁾ is likely invalid for many exposures due to violation of the constant genetic effect assumption. There are also concerns over the newly developed approach (doubly ranked method which makes the rank preserving assumption)⁷⁴⁾ as a study showed correlation of exposure with implausible outcomes such as age and sex using this method⁴³⁾. Whether these paradoxical findings were generated based on the fundamental issue with the doubly ranked method itself or dependent on the nature of the databases require additional investigations⁷⁵⁾.

Table 1 shows a brief comparison between one sample and two sample MR studies. For details, please refer to the review by Debbie Lawlor⁷⁶⁾.

Table 1 . A comparison of one sample and two sample Mendelian randomization (MR) study, with reference to the review by Lawlor¹⁾.

	One sample MR	Two sample MR
Data	– Biobank studies with instruments, exposures and outcomes – Requires large sample size to achieve sufficient statistical power	– Summary statistics of genome wide association studies (GWAS) and hence is less resource intensive – GWAS are generally large, and hence statistical power is likely not a problem
Assessment of relevance assumption	– Possible via regression of exposures on instruments in the study, alongside with F statistics and variance explained (R2) – Biased towards confounded association with weak instrument bias2)	– Can only be assessed based on calculated F statistics and R2 from summary statistics – Biased towards null with weak instrument bias2)
Assessment of independence assumption	– Assess the relation of instruments with confounders although arguably any association is likely a reflection of horizontal pleiotropy3)	– Cannot be assessed
Assessment of exclusion restriction assumption	– Assess the relation of instrument in outcome, adjusting for exposure but vulnerable to collider bias4) – Cochran’s Q – MR-Egger intercept – Assessment of pleiotropy via curated GWAS data and own data	– Cochran’s Q – MR-Egger intercept – Assessment of pleiotropy via other curated GWAS data
Assessment of non-linearity	– Possible but there are controversies surrounding the methods5)	– Not possible
Assessment of effect modification	– Able to consider effect modification through subgroup analyses	– Only possible if subgroup summary statistics are provided (e.g. ethnicity, sex)

References for the table

1. Lawlor DA. Commentary: Two-sample Mendelian randomization: opportunities and challenges. Int J Epidemiol 2016;45:908–15.

2. Pierce BL, Burgess S. Efficient design for Mendelian randomization studies: subsample and 2-sample instrumental variable estimators. Am J Epidemiol 2013;178:1177–84.

3. Carter AR, Anderson EL. Correct illustration of assumptions in Mendelian randomization. Int J Epidemiol 2024;53. doi: 10.1093/ije/dyae050

4. Au Yeung SL, Jiang C, Cheng KK, et al. Is aldehyde dehydrogenase 2 a credible genetic instrument for alcohol use in Mendelian randomization analysis in Southern Chinese men? . International Journal of Epidemiology 2013;42:318–28.

5. Hamilton FW, Hughes DA, Spiller W, et al. Non-linear Mendelian randomization: detection of biases using negative controls with a focus on BMI, Vitamin D and LDL cholesterol. Eur J Epidemiol 2024;39:451–65.

VARIATIONS OF MENDELIAN RANDOMIZATION DESIGNS

Two step Mendelian randomization (Fig. 2A)

Fig. 2 . Different forms of Mendelian randomization design, including (a) two step Mendelian randomization design; (b) bi-directional Mendelian randomization design; and (c) drug target Mendelian randomization design.

Two step MR resembles the mediation analysis in conventional observational studies, where researchers assess the relation of exposure in mediator (1^st step), and mediator in outcome (2^nd step) using MR with genetic instruments for exposure and mediator respectively⁷⁷⁾. To derive the indirect effect, product of coefficient approach can be used, which is the product of coefficient of the two steps⁶⁷⁾. However, this method may be less appropriate with multiple mediators, in which researchers may opt for simultaneous mediators adjustment via MVMR analyses to derive the direct effect. The indirect effect can then be estimated using the difference in coefficient approach, which is a subtraction of the total effect from the direct effect⁶⁷⁾.

Bi-directional Mendelian randomization (Fig. 2B)

A bi-directional MR design is intended to assess reverse causation, with two complementary sets of MR analyses. The first MR is to assess the association of exposure in outcome using genetic instruments for exposure (Forward MR). The second MR is to assess the association of outcome in exposure using genetic instruments for the outcome (Reverse MR). Any evidence of association for the reverse MR may imply the presence of reverse causation. However, if the outcome is a binary trait (e.g. disease status), the reverse MR analyses may be more difficult to interpret as any association observed is a reflection of consequence of genetic liability to the presence of binary trait rather than presence of binary trait⁷⁸⁾.

Drug target Mendelian randomization (Fig. 2C)

Drug target MR is designed to explore how genetically proxied drug target perturbation affects health outcomes so as to inform effects of the drug⁷⁹⁾. For example, statin reduces low density lipoprotein cholesterol via inhibition of the 3-hydroxy-3-methyl-glutaryl-coenzyme A reductase (HMGCR). To use genetics to infer the effect of statin, the corresponding drug target MR analyses will include variants within the HMGCR gene region (a.k.a cis-variants) which strongly predict reduced low density lipoprotein cholesterol as instruments to proxy genetic inhibition of HMGCR. Given the variants are all within the same region, these variants are likely correlated and the threshold for removing variants in linkage disequilibrium is often less stringent to guard against insufficient power (e.g. r² of 0.3) although the correlation matrix of the variants can be supplied to incorporate the correlated nature of the variants in the analyses⁸⁰⁾. Since these variants are correlated, assumptions of (sensitivity) analyses used in conventional MR analyses are likely violated although arguably any of these pleiotropic effects is likely vertical (consequence of drug target perturbation) and thus will not bias the analyses⁸¹⁾. In drug target MR, genetic colocalization (shared genetic variant between 2+ traits) is often included to rule out confounding by linkage disequilibrium in the MR analyses (assessing independence assumption)⁸²⁾. A recent review has summarized common pitfalls regarding drug target MR design, such as using actual drug use as exposure⁸³⁾. However, challenges remain with drug target MR if the drug affects multiple targets. For example, metformin affects disease outcomes via multiple pathways⁸⁴⁾. In such situations, we can only assess the individual putative effects of the drug target (e.g. AMPK, ETFDH, GPD1, and PEN2)⁸⁵⁾, but it would not be possible to derive the overall effect as the % contribution via different pathway is unknown⁸⁶⁾. As proteins are often pharmaceutical targets, MR involving proteins as exposures also follow the same instrument selection (cis variants strongly related to proteins) and analytic approach⁸²⁾.

RESOURCES CONCERNING REPORTING OF MENDELIAN RANDOMIZATION STUDIES AND GUIDELINES

Given the increase in MR studies and concerns of poor reporting⁸⁷⁾, the STROBE-MR checklist is developed to facilitate proper reporting (https://www.strobe-mr.org/)⁸⁸⁾. However, this checklist is possibly more relevant to the conventional MR design, where not all items are relevant to other forms of MR designs, such as drug target MR⁸⁹⁾. Nevertheless, there are various reviews to help guide proper design of corresponding MR studies, such as the ones related to drug target MR^79,82), or intrauterine exposures⁹⁰⁾. There are also reviews on MR^91,92), as well as other online resources (e.g. MR-dictionary (https://mr-dictionary.mrcieu.ac.uk/)) which researchers may find useful.

EXAMPLES OF TWO SAMPLE MR STUDY AND DRUG TARGET MR STUDY

Example 1: Iwagami M et al., Blood Lipids and the Risk of Colorectal Cancer: Mendelian Randomization Analyses in the Japanese Consortium of Genetic Epidemiology Studies. Cancer Prevention and Research. 2022: 15: 827–836

This was a two sample MR study exploring the association of lipids (exposure) in colorectal cancer risk (outcome) using East Asian data⁹³⁾. For the instrument-exposure associations, the authors identified strong and independent genetic predictors of lipids (across the entire genome, p value < 5 × 10⁻⁸) from the East Asian specific analyses of the lipid GWAS although the corresponding instrument-exposure associations were derived based on a meta-analysis of three Japanese GWAS (regardless of statistical significance). For the instrument-outcome associations, the authors extracted the information based on meta-analyses of five individual Japanese GWAS, and GWAS summary statistics from Biobank Japan. Inverse variance weighed method was used as main analysis, with several sensitivity analyses to assess robustness of findings, such as MR-Egger, weighted median, weighted mode, MR-PRESSO, as well as employing different approaches in obtaining the instrument-exposure associations in the Japanese cohorts. The main finding from this study was that higher total cholesterol was associated with higher risk of colorectal cancer although evidence for other lipid traits was less conclusive.

Example 2: Yarmonlinksy et al., Association Between Genetically Proxied Inhibition of HMG-CoA Reductase and Epithelial Ovarian Cancer. JAMA. 2020. 323: 646–655

This was a two sample MR study exploring the association of genetically proxied inhibition of HMGCR with epithelial ovarian cancer⁹⁴⁾. Variants to proxy HMGCR inhibition were identified based on a search of weakly correlated variants (r² < 0.2) within HMGCR (±100 kb of HMGCR (Entrez Gene: 3156)) strongly associated with low density lipoprotein cholesterol in a lipid GWAS. These instruments were applied to GWAS summary statistics of ovarian cancer in the general population and those who had BRCA1/2 mutations. Main analysis was inverse variance weighted method incorporating weak linkage disequilibrium. Sensitivity analyses included colocalization, leave-one-out analysis, and multivariable MR adjusting for ovarian cancer risk factors associated with HMGCR variants. The main finding from this study was that genetically proxied HMGCR inhibition was associated with lower risk of ovarian cancer. Drug target MR analyses for other genetically proxied targets of lipid modifying medications and conventional MR analyses of low density lipoprotein cholesterol showed no association.

EMERGING CONCERNS OVER MENDELIAN RANDOMIZATION STUDIES

With the increasing availability of user written packages and publicly available genetic databases and biobanks, there is an explosion of MR studies in the past years²⁹⁾. However, this also raises substantial concerns over the quality of these studies leading to credibility crisis⁹⁵⁾. For example, there are several MR studies using air pollution as exposures^96,97). Unfortunately, given air pollution level is derived based on residential addresses, any genetic associations with air pollution level is likely reflection of severe horizontal pleiotropy (e.g. via determinants of where participants lived) and makes the corresponding results challenging to interpret⁹⁸⁾. Others considered these MR studies are conducted simply because data is available rather than motivated by meaningful research gaps⁹⁹⁾. Some commented that this method is being hijacked by paper mills business given its ease to conduct such studies²⁹⁾. From the editors and reviewers’ perspective, request to handle/review MR studies will be increasing and having a good understanding of the issues surrounding this design could help facilitate the peer review processes and eventual disposition of impactful MR studies.

CONCLUSION

In conclusion, this paper introduces concepts and issues surrounding MR studies. A properly conducted MR study can provide strong, alternative evidence to evaluate causality of risk factors in non-communicable diseases and infectious diseases^100,101). However, as with all epidemiologic designs, a clear understanding of the research question is crucial to evaluate the appropriateness of MR studies to address the question at hand so as to maximize the positive impact of MR in shaping the evidence base for clinical practice and public health policies.

CONFLICT OF INTEREST

SLAY is currently an advisory board member of Annals of Clinical Epidemiology. He also received honoraria from Standard BioTools for scientific presentations on proteomic studies that was unrelated to this seminar paper. Other authors declared no other conflict of interest.

FUNDING

AG and MI were supported for this study by the National Cancer Center Research and Development Fund (2022-A-20). SLAY was supported by the Health and Medical Research Fund Research Fellowship Scheme (01150037), Health Bureau, HKSAR, China for training in Mendelian randomization studies. The funders had no role in the design, analyses, interpretation of results or writing of the paper.

AUTHORS’ CONTRIBUTIONS

SLAY wrote the first draft of this seminar paper, with feedback from SL, MI, and AG. All authors read and approved the final manuscript. SLAY had primary responsibility for the final content.

DISCLAIMER

Shiu Lun Au Yeung is one of the Editorial Board members of Annals of Clinical Epidemiology. This author was not involved in the peer-review or decision-making process for this paper.