Published | Last updated
What Is Mendelian Randomization?
Mendelian randomization is a way of using inherited genetic variation to ask whether a modifiable exposure, such as LDL cholesterol, causes an outcome such as coronary artery disease. It builds on Mendel’s laws of inheritance and on instrumental variable estimation, which makes it possible to estimate causal effects even when some confounders were never measured.5 The original argument was that the random assortment of genes from parents to offspring at gamete formation and conception means a polymorphism that mimics the biological link with an exposure is not generally susceptible to the reverse causation or confounding that distorts conventional observational studies.1
In simple words: if a variant that nudges an exposure such as LDL cholesterol is assigned by chance, comparing people with and without that variant is like comparing two groups from a coin flip, provided the variant influences the outcome only through that exposure.
Every analysis is built from three linked pieces.
- The instrument. One or more genetic variants associated with the exposure.
- The exposure. The modifiable risk factor under study, for example LDL cholesterol.
- The outcome. The disease or trait of interest, for example coronary artery disease.
The method has limits that an early review already named: confounding by polymorphisms in linkage disequilibrium with the variant, polymorphisms with several phenotypic effects, a lack of suitable polymorphisms for many exposures, and canalization, the buffering of genetic effects during development.1
The Three Core Assumptions Behind Every Mendelian Randomization Study
STROBE-MR item 5 asks authors to state the three core instrumental variable assumptions explicitly for the main analysis, together with the assumptions of any additional or sensitivity analysis. Item 7 asks them to describe the methods or prior knowledge used to assess those assumptions.3
Swipe sideways to see every column.
| Assumption | In plain terms | How it is probed |
|---|---|---|
| Relevance | The variant is associated with the exposure. | Report the variance explained (R²) and, particularly in a one-sample setting, the F statistic as a measure of instrument strength.6 |
| Independence | The variant is not associated with the outcome through confounding pathways. | Validity relies on independence from confounders at population level; within-family analysis is an option where data and power allow.6 |
| Exclusion restriction | The variant does not affect the outcome except potentially through the exposure. | Pleiotropy checks: MR-Egger, weighted median, heterogeneity statistics and outlier methods, covered below. |
The assumptions are the same ones Burgess and colleagues list in their guidelines: the variant is associated with the exposure, it is not associated with the outcome due to confounding pathways, and it does not affect the outcome except potentially via the exposure.6
Weak instruments
A weak instrument biases the estimate in the direction of the confounded observational association, and the size of the bias depends on the F statistic. Burgess and Thompson also showed that the commonly cited rule of thumb that F above 10 avoids bias is misleading.7 In simulations with uncorrelated variants and summarized data, Burgess and colleagues found weak instrument bias was small when each variant had a P value below 0.00001 for its association with the risk factor.8 Our companion article on instrumental variable approaches covers the general logic outside genetics.
One-Sample, Two-Sample and Multivariable Designs, and the Estimators
In a one-sample analysis the data on the risk factor and the outcome come from a single source. In a two-sample analysis they come from non-overlapping datasets.9 Public summarized data make a two-sample analysis transparent, because a third party with the same data can reproduce it, while individual-level data in a one-sample setting are required to investigate non-linear effects.6 Weak-instrument bias also differs: in a one-sample setting it is toward the confounded association, and in a two-sample setting without overlap it is toward the null.6
Several estimators are usually run side by side.
- Inverse-variance weighted (IVW) is the most commonly used method with multiple variants. It is the most efficient, but is biased if the average pleiotropic effect differs from zero.6 It can be obtained from summarized data as an inverse-variance weighted average of the variant-specific associations.8
- MR-Egger regression takes the slope as the causal estimate and the intercept as the average pleiotropic effect. It allows every variant to be pleiotropic but needs the InSIDE assumption, and it is often less efficient.10,6
- The weighted median is consistent even when up to 50% of the information comes from invalid instruments.4
- Multivariable MR uses variants associated with several related exposures to estimate the effect of each exposure on one outcome, and can be used to estimate mediation within the MR framework.11
For a simulation that sets these estimators against different kinds of pleiotropy, see Editverse’s article on what each of the four estimators actually protects you from.
A Real Case: HDL and LDL Cholesterol and Coronary Artery Disease
Bowden and colleagues published an open-access paper (CC BY 4.0) that tested the weighted median against IVW and MR-Egger on summary data for 185 genetic variants from Do and colleagues, using lipid associations from the Global Lipids Genetics Consortium and coronary artery disease associations from the CARDIoGRAM consortium.4 Estimates are log odds ratios per 1 standard deviation higher lipid. We took the estimates and standard errors from their Table 4 (the column for all variants associated with the exposure at genome-wide significance) and recomputed odds ratios, 95% intervals and p-values in R using a normal approximation.4

Swipe sideways to see every column.
| Estimator | LDL cholesterol, OR (95% CI) | HDL cholesterol, OR (95% CI) |
|---|---|---|
| Inverse-variance weighted | 1.62 (1.44 to 1.82) | 0.78 (0.68 to 0.89) |
| Weighted median | 1.58 (1.39 to 1.80) | 0.93 (0.81 to 1.07) |
| MR-Egger slope | 1.85 (1.51 to 2.27) | 0.99 (0.79 to 1.24) |
| MR-Egger intercept (SE), computed p | -0.009 (0.005), p = 0.072 | -0.014 (0.005), p = 0.005 |
LDL cholesterol gave a consistent answer across all three estimators, with no strong sign of directional pleiotropy (MR-Egger intercept p = 0.072). For HDL cholesterol the IVW interval excluded 1, while the weighted median and MR-Egger intervals included 1 and the MR-Egger intercept differed from zero (p = 0.005). The authors conclude that the IVW causal effect of HDL cholesterol is suspect.4 When they restricted the analysis to variants whose strongest association was with HDL cholesterol, the pattern held: IVW OR 0.87 (0.77 to 0.99), weighted median OR 0.94 (0.82 to 1.06), MR-Egger slope OR 1.10 (0.89 to 1.35) (computed by Editverse).4
What this case shows: an IVW-only report would have presented HDL cholesterol as protective, and the extra estimators changed the reading. What it does not show: that HDL cholesterol has no effect, because the weighted median and MR-Egger intervals are wide; or that the estimators are always right. We redrew published numbers and did not re-run the 185-variant data, which we did not obtain.
Pleiotropy, the Central Threat to Validity
Horizontal pleiotropy, where a variant reaches the outcome by a pathway that bypasses the exposure, breaks exclusion restriction. Verbanck and colleagues developed MR-PRESSO to identify horizontal pleiotropic outliers in multi-instrument summary-level MR, and applied it to complex traits and diseases. They found horizontal pleiotropy detectable in over 48% of significant causal relationships, with distortions of causal estimates ranging on average from -131% to 201%, and false-positive causal relationships in up to 10% of relationships.12 Table 1 of the Burgess guidelines notes that MR-PRESSO has a very high false positive rate when several instruments are invalid, so it is one check among several.6
A defensible analysis runs several checks together.
- MR-Egger intercept and the slope as a sensitivity estimate.10
- Weighted median and other robust estimators.4
- Heterogeneity. The guidelines recommend reporting Cochran’s Q or the I² statistic for a polygenic analysis.6
- Leave-one-out and Steiger filtering. STROBE-MR lists leave-one-out plots among additional plots,3 and Steiger filtering removes variants whose association with the outcome is stronger than with the exposure.6
- A liberal and a conservative variant set. Pleiotropy is generally more likely to bias estimates away from the null, so a null finding in a liberal analysis is more convincing evidence of a true null.6
Sample Overlap and Weak-Instrument Bias
Many large consortia share participants, so the exposure and outcome datasets in a two-sample analysis can overlap. Burgess and colleagues simulated this for a continuous outcome and a null causal effect: a one-sample analysis with a relative bias of 10% (corresponding to an F parameter of 10) fell to 5% with 50% overlap and 3% with 30% overlap, because bias is a linear function of the overlap proportion.9 STROBE-MR item 10d asks authors of two-sample studies to report the number of individuals who overlap between the exposure and outcome studies.3

More Published Examples
LDL cholesterol and coronary heart disease
Ference and colleagues combined Mendelian randomization studies of 9 polymorphisms in 6 genes, with no evidence of heterogeneity. In non-overlapping data from 312,321 participants, naturally random allocation to long-term lower LDL cholesterol was associated with a 54.5% (95% CI 48.8% to 59.5%) reduction in coronary heart disease risk per mmol/l lower LDL cholesterol, about 3-fold the reduction seen during statin treatment started later in life.13 Using only five variants and summarized data, Burgess and colleagues estimated that a 30% reduction in LDL cholesterol lowers coronary artery disease risk by 67% (95% CI 54% to 76%).8
Drug-target Mendelian randomization
Gagnon and colleagues genotyped and RNA-sequenced 504 human liver samples and identified a variant (rs472495) explaining 5.6% of liver PCSK9 expression to mimic RNA interference, and used PCSK9 R46L to model antibody-based inhibition. For each standard deviation decrease in apolipoprotein B, the odds ratios for coronary artery disease were 0.40 (95% CI 0.31 to 0.51) and 0.48 (0.43 to 0.55) respectively.14
Vitamin D
Fang and colleagues pooled 133 Mendelian randomization publications covering 275 outcomes up to September 2022. Genetically high 25-hydroxyvitamin D was associated with lower risk of multiple sclerosis, among other outcomes, and with higher risk of some, such as kidney stone disease. The multiple sclerosis association was consistent across instrument sets, while most others appeared only with variants not limited to the vitamin D pathway, and study quality was heterogeneous.15 Zhao and colleagues used 332,984 UK Biobank participants, 23,089 with at least one autoimmune disease, and reported an odds ratio of 0.95 (95% CI 0.91 to 0.99) per 10 ng/ml higher 25-hydroxycholecalciferol for the autoinflammation group but not the autoimmunity group.16
Alcohol and vascular disease in China
Millwood and colleagues genotyped 161,498 participants of the China Kadoorie Biobank for ALDH2-rs671 and ADH1B-rs1229984. Self-reported intake had U-shaped associations with stroke and myocardial infarction in men, but genotype-predicted intake did not. The relative risk per 280 g per week was 1.58 (1.36 to 1.84) for intracerebral haemorrhage, 1.27 (1.13 to 1.43) for ischaemic stroke and 0.96 (0.78 to 1.18) for myocardial infarction. The authors concluded that the apparently protective effects of moderate alcohol intake against stroke are largely non-causal.17
When Should You Use Mendelian Randomization?
- A randomized trial is not feasible. The method is most useful when randomized trials cannot be done and observational studies are biased by confounding or reverse causality.2
- You need to probe a drug target. MR has been applied to predict the efficacy and safety of drugs targeting cardiovascular risk factors and to explore repurposing.18
- Adequately sized summarized data exist. Public GWAS summary data allow a two-sample design without individual genotypes.6
The method fits poorly when no genetic variants can be justified as instruments for the exposure, because the relevance assumption then fails, or when the available data are too small, because MR studies often have low power.6 If you are choosing between a genetic design and a survey or registry analysis, see which health database your study should use.
How to Conduct a Mendelian Randomization Study
- 1State the causal hypothesis. STROBE-MR items 2 and 3 ask for the rationale, why MR is a helpful method, and specific objectives including prespecified causal hypotheses.3
- 2Choose the design and data. Decide on one-sample or two-sample, and individual-level or summarized data.6
- 3Select the variants. A common approach is to include variants associated with the exposure at genome-wide significance, typically P below 5 × 10-8. Selecting in the same dataset as the analysis risks winner’s curse and worsens weak instrument bias.6
- 4Harmonize alleles. Align exposure and outcome alleles. For palindromic variants (A/T or C/G) with allele frequency near 50%, analysts may drop the variant if the orientation cannot be verified.6
- 5Run the primary and sensitivity analyses. IVW as the primary estimate for multiple variants, then robust estimators, heterogeneity statistics and plots.6
- 6Report to STROBE-MR. The checklist has 20 main items and 30 subitems, including software and package versions (item 9a), pre-registration status (item 9b), and data and code sharing (item 19).19,3
The STROBE-MR statement extends the general logic of the STROBE reporting guideline.
Data Sources and Software for MR Analysis
Descriptions below are taken from each resource’s own page, as read on 4 November 2026.
Swipe sideways to see every column.
| Resource | What it is | Typical use |
|---|---|---|
| OpenGWAS (MRC IEU)20 | A database of more than 50,000 GWAS summary datasets, for querying or download. | Exposure and outcome summary statistics for two-sample designs. |
| FinnGen21 | A project that has collected and analysed genome and health data from 500,000 Finnish biobank donors. | Disease-outcome genetic associations. |
| NHGRI-EBI GWAS Catalog22 | The NHGRI-EBI Catalog of human genome-wide association studies. Its summary statistics section offers documentation and access to full summary statistics for studies where available. | Finding and cross-checking variant-trait associations. |
| TwoSampleMR (R package)23 | An R package for MR with GWAS summary data that can obtain data from OpenGWAS. | Harmonization, IVW, MR-Egger, weighted median. |
| MendelianRandomization (R package)24 | A CRAN package that encodes several MR methods for summarized data. | An alternative implementation of the core estimators. |
| mrpresso (R package)6 | Listed in the Burgess guidelines as the software for MR-PRESSO. | Outlier detection; see the caution above. |
Common Mistakes in Mendelian Randomization
- Not reporting instrument strength. Report R² and the F statistic, and remember that the F above 10 rule of thumb is misleading.6,7
- Reporting only the IVW estimate. The HDL cholesterol case above shows how a single estimator can mislead.4
- Ignoring sample overlap. Report the number of overlapping individuals and expect bias to vary with the degree of overlap.9,3
- Mixing populations without care. If the two samples represent different ethnic groups, patterns of linkage disequilibrium can differ, so a variant may be less strongly associated with the exposure in the outcome dataset.6
- Reading a wide null interval as no effect. A null in the HDL case came with intervals from about 0.8 to 1.2, which are compatible with modest effects in either direction.4
- Using causal language loosely. STROBE-MR item 16 asks authors to use causal language carefully and to clarify that estimates may give causal effects only under certain assumptions.3
Mendelian Randomization vs. Other Causal Inference Methods
The table states the defining feature of each design for the two problems MR is meant to address.
Swipe sideways to see every column.
| Method | Confounding | Reverse causation |
|---|---|---|
| Randomized controlled trial | Random assignment of the intervention. | By design, the intervention is assigned before outcomes are measured. |
| Mendelian randomization | Reduced if the variant is independent of confounders.6 | Genotype is fixed at conception, so it is not generally susceptible.1 |
| Propensity score matching | Balances measured covariates only. | Not addressed by matching on baseline covariates. |
| Multivariable regression | Adjusts for measured covariates only. | Not addressed by covariate adjustment. |
Because each design has different key sources of bias, triangulation across approaches strengthens confidence when they agree, and MR can serve as one strand of it.25,5 For related background see our guides to propensity score matching and causal inference in epidemiology.
Need the Analysis Run or Checked? Consult Editverse
Editverse subject-matter experts work to COPE, ICMJE and GPP 2022 guidance, with institutional and country-specific ethics followed.
Editverse
Mendelian Randomization Study
Need a full two-sample MR study built to journal standards?
You send: your exposure, your outcome and the causal question. Summary statistics are public; we obtain them.
You get: a two-sample MR study with instrument strength quantified, Steiger filtering, four pleiotropy-robust estimators, MR-PRESSO outlier detection, instrument PheWAS and independent replication, plus a complete manuscript, scatter and funnel plots, runnable TwoSampleMR code and a STROBE-MR checklist.
When: You want the whole study run and written up. 15 business days.
Mendelian Randomization Study · $2,900
Editverse
Feasibility Check
Not sure an MR design can support your question?
You send: your research question.
You get: a written recommendation on whether the question is already answered, whether the data can support it and whether you have the power, including instrument strength for a Mendelian randomization design. Credited in full against any study you commission within six months.
When: Before you commit to a study. 5 business days.
Feasibility & Protocol Assessment · $450
Editverse
Statistical Analysis
Have your data and need the analysis run in R?
You send: your data, research questions and preferred software.
You get: statistical analysis using your preferred software (SPSS, R, STATA or SAS), with annotated code, complete output files and an interpretation guide.
When: You need publication-ready results in 4-6 business days.
SPSS/R/STATA Statistical Analysis · $400
Free Research Tools from Editverse
- Sample Size and Power Calculator: power analysis for the study you are planning.
- Statistical Test Selector: the test, its assumptions and what to report.
- Free Research Tools for Medical Researchers: the full set of browser-based tools.
Frequently Asked Questions About Mendelian Randomization
Is Mendelian randomization the same as a randomized controlled trial?
No. Mendelian randomization works with observational data and treats inherited genetic variants as instruments. Alleles are passed from parents to offspring by random segregation, which resembles random allocation of a treatment in a trial and should lessen confounding. The estimate also rests on assumptions that a trial does not need, so it complements trial evidence and does not replace it.
What are the three core assumptions of Mendelian randomization?
Relevance, independence and exclusion restriction. In plain terms, the variant is associated with the exposure, it is not associated with the outcome through confounding pathways, and it does not affect the outcome except potentially through the exposure. STROBE-MR item 5 asks authors to state these explicitly.
Do I need individual-level genotype data to run Mendelian randomization?
Not for a two-sample analysis, which can use published summarized associations of each variant with the exposure and with the outcome. Individual-level data in a one-sample setting are required to investigate non-linear effects. A practical advantage of public summarized data is that a third party can reproduce the analysis.
What is the difference between one-sample and two-sample Mendelian randomization?
In a one-sample analysis the data on the risk factor and outcome come from a single source. In a two-sample analysis they come from non-overlapping datasets. The direction of weak-instrument bias differs: toward the confounded association in a one-sample setting, and toward the null in a two-sample setting without sample overlap.
Why report more than the inverse-variance weighted estimate?
Because the inverse-variance weighted method gives a consistent estimate only if all variants are valid or pleiotropy is balanced. In the published lipid example above, the IVW odds ratio for HDL cholesterol was 0.78, while the weighted median gave 0.93 and MR-Egger 0.99 with intervals that include 1. Reporting several estimators shows readers whether the result depends on one method.
What does the MR-Egger intercept tell me?
In MR-Egger regression the intercept estimates the average pleiotropic effect across the variants, and the slope gives the causal estimate. An intercept that differs from zero points to directional pleiotropy. The method needs the InSIDE assumption, that pleiotropic effects are independent of the variant-exposure associations, and it is often less efficient than other estimators.
What does STROBE-MR require?
STROBE-MR is a reporting guideline with 20 main items and 30 subitems. Examples include stating the three instrumental variable assumptions (item 5), describing how assumptions were assessed (item 7), reporting sensitivity analyses (items 8 and 13) and, for two-sample designs, the number of individuals who overlap between the exposure and outcome studies (item 10d).
Can Mendelian randomization prove causation?
Not on its own. Mendelian randomization tests a causal hypothesis under specific assumptions, and some of those assumptions cannot be fully verified. Use causal language carefully, report the sensitivity analyses, and treat the result as one line of evidence in a triangulation with other designs.
References
- Davey Smith G, Ebrahim S. ‘Mendelian randomization’: can genetic epidemiology contribute to understanding environmental determinants of disease? Int J Epidemiol. 2003;32(1):1-22. PMID 12689998. doi.org/10.1093/ije/dyg070
- Sekula P, Del Greco M F, Pattaro C, Kottgen A. Mendelian randomization as an approach to assess causality using observational data. J Am Soc Nephrol. 2016;27(11):3253-3265. PMID 27486138. doi.org/10.1681/ASN.2016010098
- Skrivankova VW, Richmond RC, Woolf BAR, et al. Strengthening the reporting of observational studies in epidemiology using mendelian randomisation (STROBE-MR): explanation and elaboration. BMJ. 2021;375:n2233. Open access, CC BY 4.0. doi.org/10.1136/bmj.n2233
- Bowden J, Davey Smith G, Haycock PC, Burgess S. Consistent estimation in Mendelian randomization with some invalid instruments using a weighted median estimator. Genet Epidemiol. 2016;40(4):304-314. Open access, CC BY 4.0; odds ratios, intervals and p-values recomputed by Editverse in R from Table 4. doi.org/10.1002/gepi.21965
- Sanderson E, Glymour MM, Holmes MV, et al. Mendelian randomization. Nat Rev Methods Primers. 2022;2:6. PMID 37325194. doi.org/10.1038/s43586-021-00092-5
- Burgess S, Davey Smith G, Davies NM, et al. Guidelines for performing Mendelian randomization investigations: update for summer 2023. Wellcome Open Res. 2019;4:186 (version 3). Open access, CC BY 4.0. doi.org/10.12688/wellcomeopenres.15555.3
- Burgess S, Thompson SG. Avoiding bias from weak instruments in Mendelian randomization studies. Int J Epidemiol. 2011;40(3):755-764. PMID 21414999. doi.org/10.1093/ije/dyr036
- Burgess S, Butterworth A, Thompson SG. Mendelian randomization analysis with multiple genetic variants using summarized data. Genet Epidemiol. 2013;37(7):658-665. PMID 24114802. doi.org/10.1002/gepi.21758
- Burgess S, Davies NM, Thompson SG. Bias due to participant overlap in two-sample Mendelian randomization. Genet Epidemiol. 2016;40(7):597-608. Open access, CC BY 4.0; linear relation checked by Editverse in R. doi.org/10.1002/gepi.21998
- Bowden J, Davey Smith G, Burgess S. Mendelian randomization with invalid instruments: effect estimation and bias detection through Egger regression. Int J Epidemiol. 2015;44(2):512-525. PMID 26050253. doi.org/10.1093/ije/dyv080
- Sanderson E. Multivariable Mendelian randomization and mediation. Cold Spring Harb Perspect Med. 2021;11(2):a038984. PMID 32341063. doi.org/10.1101/cshperspect.a038984
- Verbanck M, Chen CY, Neale B, Do R. Detection of widespread horizontal pleiotropy in causal relationships inferred from Mendelian randomization between complex traits and diseases. Nat Genet. 2018;50(5):693-698. PMID 29686387. doi.org/10.1038/s41588-018-0099-7
- Ference BA, Yoo W, Alesh I, et al. Effect of long-term exposure to lower low-density lipoprotein cholesterol beginning early in life on the risk of coronary heart disease: a Mendelian randomization analysis. J Am Coll Cardiol. 2012;60(25):2631-2639. PMID 23083789. doi.org/10.1016/j.jacc.2012.09.017
- Gagnon E, Gill D, Bourgault J, et al. RNA interference versus antibody-based PCSK9 inhibition for the prevention of cardiovascular disease: a drug-target Mendelian randomization study. Cardiovasc Res. 2025;121(7):1066-1075. PMID 40347490. doi.org/10.1093/cvr/cvaf078
- Fang A, Zhao Y, Yang P, Zhang X, Giovannucci EL. Vitamin D and human health: evidence from Mendelian randomization studies. Eur J Epidemiol. 2024;39(5):467-490. PMID 38214845. doi.org/10.1007/s10654-023-01075-4
- Zhao SS, Mason A, Gjekmarkaj E, Yanaoka H, Burgess S. Associations between vitamin D and autoimmune diseases: Mendelian randomization analysis. Semin Arthritis Rheum. 2023;62:152238. PMID 37437450. doi.org/10.1016/j.semarthrit.2023.152238
- Millwood IY, Walters RG, Mei XW, et al. Conventional and genetic evidence on alcohol and vascular disease aetiology: a prospective study of 500 000 men and women in China. Lancet. 2019;393(10183):1831-1842. Open access, CC BY 4.0. doi.org/10.1016/S0140-6736(18)31772-0
- Larsson SC, Butterworth AS, Burgess S. Mendelian randomization for cardiovascular diseases: principles and applications. Eur Heart J. 2023;44(47):4913-4924. PMID 37935836. doi.org/10.1093/eurheartj/ehad736
- Skrivankova VW, Richmond RC, Woolf BAR, et al. Strengthening the Reporting of Observational Studies in Epidemiology Using Mendelian Randomization: the STROBE-MR statement. JAMA. 2021;326(16):1614-1621. PMID 34698778. doi.org/10.1001/jama.2021.18236
- MRC Integrative Epidemiology Unit. OpenGWAS (home page, read 4 November 2026). gwas.mrcieu.ac.uk/
- FinnGen. Project home page (read 4 November 2026). www.finngen.fi/en
- NHGRI-EBI GWAS Catalog. Home page (read 4 November 2026). www.ebi.ac.uk/gwas/
- MRC Integrative Epidemiology Unit. TwoSampleMR documentation (version 0.7.13 shown on 4 November 2026). mrcieu.github.io/TwoSampleMR/
- CRAN. MendelianRandomization package page (version 0.10.0, published 2024-04-12, read 4 November 2026). cran.r-project.org/package=MendelianRandomization
- Lawlor DA, Tilling K, Davey Smith G. Triangulation in aetiological epidemiology. Int J Epidemiol. 2016;45(6):1866-1886. PMID 28108528. doi.org/10.1093/ije/dyw314
Related Reading
Editverse Research Team | Updated November 2026
