Neil M DaviespersName, Michael V. HolmespersName, George Davey SmithpersName (2018date). “Reading Mendelian randomisation studies: a guide, glossary, and checklist for clinicians.” BMJtitle·venue.
Object of study — A METHOD WHOSE VALIDITY IS THREE NAMED ASSUMPTIONS, AND A TERM WHOSE TWO MODIFIERS DECIDE WHETHER A STUDY IS RUINED OR WORKING - Mendelian randomisationterm, read through its own glossary and a critical-appraisal checklist written for clinicians
Research question — how a clinician should read a Mendelian randomisationterm study: which assumptions have to hold, which reported quantities test them, and what the estimate actually refers to
Contribution — a glossary of the method's terms, a table of the three instrumental-variable assumptions with the tools that assess each, and a critical appraisal checklist covering assumptions, methods reporting, data presentation, interpretation and clinical implications
phenomenon · claim-type · cross-domain-transfer · mechanism · claim-relation · replication-status · limitation — grounded in the full text| phenomenon | causal inference from observational data, where the thing that makes the inference possible - a genetic variant fixed at conception - also fixes the estimand to a lifetime of exposure rather than an intervention — “But this must be approached with care because genetic variants generally relate to lifelong differences in a risk factor, not the effects at a specific time. ALDH2 is associated with alcohol consumption from adolescence through middle age, 76 so the Mendelian randomisation results (coupled with biological knowledge) indicate that consuming less alcohol would lead to lower blood pressure on average, but do not necessarily inform the effect size of an intervention at a specific time in life for a specific duration.” | |
| claim-type | method — “Mendelian randomisation is a research method that provides evidence about putative causal relations between modifiable risk factors and disease, using genetic variants as natural experiments Mendelian randomisation is less likely to be affected by confounding or reverse causation than conventional observational studies Like all analytical approaches, however, Mendelian randomisation depends on assumptions, and the plausibility of these assumptions must be assessed Moreover, the relevance of the results for clinical decisions should be interpreted in light of other sources of evidence We provide a critical appraisal checklist that can be used to assess and interpret Mendelian randomisation studies” | |
| method — evidence | — “Valid instrumental variables are defined by three key assumptions ( table 1 , fig 1 ): that they associate with the risk factor of interest (the relevance assumption); that they share no common cause with the outcome (the independence assumption); and that they do not affect the outcome except through the risk factor (the exclusion restriction assumption).” | |
| object of study — evidence | — “Valid instrumental variables are defined by three key assumptions ( table 1 , fig 1 ): that they associate with the risk factor of interest (the relevance assumption); that they share no common cause with the outcome (the independence assumption); and that they do not affect the outcome except through the risk factor (the exclusion restriction assumption).” | |
| research question — evidence | — “Mendelian randomisation uses genetic variation as a natural experiment to investigate the causal relations between potentially modifiable risk factors and health outcomes in observational data. As with all epidemiological approaches, findings from Mendelian randomisation studies depend on specific assumptions. We provide explanations of the information typically reported in Mendelian randomisation studies that can be used to assess the plausibility of these assumptions and guidance on how to interpret findings from Mendelian randomisation studies in the context of other sources of evidence” | |
| contribution — evidence | — “Mendelian randomisation is a research method that provides evidence about putative causal relations between modifiable risk factors and disease, using genetic variants as natural experiments Mendelian randomisation is less likely to be affected by confounding or reverse causation than conventional observational studies Like all analytical approaches, however, Mendelian randomisation depends on assumptions, and the plausibility of these assumptions must be assessed Moreover, the relevance of the results for clinical decisions should be interpreted in light of other sources of evidence We provide a critical appraisal checklist that can be used to assess and interpret Mendelian randomisation studies” | |
| transfer | from econometrics (instrumental variables): the instrumental-variable estimator is imported from econometrics and its instrument is a genetic variant rather than an economic one — “Mendelian randomisation is an analytical method that uses genetic variants as instrumental variables for modifiable risk factors that affect population health.” | transfers[0] |
| transfer | from economics (the natural experiment): genetic variation is read as a natural experiment, the design economics uses to get causal estimates out of observational data — “Mendelian randomisation uses genetic variation as a natural experiment to investigate the causal relations between potentially modifiable risk factors and health outcomes in observational data.” | transfers[1] |
| mechanism | ONE TERM, TWO MODIFIERS, AND THE DIFFERENCE BETWEEN A FATAL FLAW AND THE METHOD ITSELF: `Pleiotropic effects - the effects of a genetic variant on multiple biological pathways. These can either affect the outcome through another trait or pathway to the one under investigation, known as HORIZONTAL PLEIOTROPY, or affect other traits through the risk factor of interest, known as VERTICAL PLEIOTROPY. HORIZONTAL PLEIOTROPY IS A violation of the instrumental variable assumptions ...; this is problematic for Mendelian randomisation studies. Vertical pleiotropy is in fact the essence of Mendelian randomisation ... and is in general NOT PROBLEMATIC.` The two differ only in which arrow the second effect travels along — “Pleiotropic effects —the effects of a genetic variant on multiple biological pathways. These can either affect the outcome through another trait or pathway to the one under investigation, known as horizontal pleiotropy, or affect other traits through the risk factor of interest, known as vertical pleiotropy. Horizontal pleiotropy is a violation of the instrumental variable assumptions because the effects of the genetic variant on the outcome are not exclusively through the risk factor; this is problematic for Mendelian randomisation studies. Vertical pleiotropy is in fact the essence of Mendelian randomisation—showing that one factor influences a downstream outcome—and is in general not problematic 1” | mechanisms[0] |
| mechanism | AND THE VALIDITY CONDITIONS ARE THREE, NAMED, AND EACH HAS ITS OWN TEST: `Valid instrumental variables are defined by three key assumptions ...: that they associate with the risk factor of interest (the relevance assumption); that they share no common cause with the outcome (the independence assumption); and that they do not affect the outcome except through the risk factor (the exclusion restriction assumption).` And they are per-variant: `Generally, the three key assumptions must hold for each of the genetic variants` - so an instrument built from a hundred variants makes a hundred assumptions — “Valid instrumental variables are defined by three key assumptions ( table 1 , fig 1 ): that they associate with the risk factor of interest (the relevance assumption); that they share no common cause with the outcome (the independence assumption); and that they do not affect the outcome except through the risk factor (the exclusion restriction assumption).” | mechanisms[1] |
| mechanism | AND THE PROBLEM IT EXISTS TO SOLVE IS NAMED AS UNVERIFIABLE: conventional adjustment works `if we measure enough confounders ... But this assumption is unverifiable; if it does not hold, then findings from observational studies will be biased estimates of causal effects ... Measuring a confounder does not perfectly characterise it, so measurement error leads to residual confounding, even after apparent statistical adjustment.` The adjustment that looks complete is the thing being replaced, and the reason is measurement error - which is althubaiti2016's subject, tagged an hour ago in the same discipline — “Many epidemiological methods attempt to correct for, or minimise, observed differences in confounders between study participants. These methods can give useful evidence about causal relations if we measure enough confounders so that, after adjustment or matching, study participants who consume different amounts of alcohol are otherwise comparable. But this assumption is unverifiable; if it does not hold, then findings from observational studies will be biasedestimates of causal effects. People who consume more alcohol may also have other risk factors for cardiovascular disease, such as smoking more heavily than those with lower alcohol consumption. The confounding factor (smoking) induces a positive association between the risk factor (alcohol) and an outcome (blood pressure); interpreting this as causal would be misleading. Measuring a confounder does not perfectly characterise it, so measurement error leads to residual confounding, even after apparent statistical adjustment.” | mechanisms[2] |
| mechanism | AND THE ESTIMAND IS NOT THE ONE A CLINICIAN WANTS: `genetic variants generally relate to lifelong differences in a risk factor, not the effects at a specific time ... the Mendelian randomisation results ... indicate that consuming less alcohol would lead to lower blood pressure on average, but do not necessarily inform the effect size of an intervention at a specific time in life for a specific duration.` The checklist says it in one line: `the numerical effect estimates may not be clinically meaningful`. A method that answers `does this cause that` and not `what happens if I change it now` — “But this must be approached with care because genetic variants generally relate to lifelong differences in a risk factor, not the effects at a specific time. ALDH2 is associated with alcohol consumption from adolescence through middle age, 76 so the Mendelian randomisation results (coupled with biological knowledge) indicate that consuming less alcohol would lead to lower blood pressure on average, but do not necessarily inform the effect size of an intervention at a specific time in life for a specific duration.” | mechanisms[3] |
| mechanism | AND A WORKED CASE WHERE THE INSTRUMENT CANNOT ANSWER THE QUESTION IT LOOKS LIKE IT ANSWERS: `Variants in the CHRNA5 gene that associate with increased heaviness of smoking are ... associated with increased risk of lung cancer. People with these variants are also likely to find it more difficult to reduce smoking after diagnosis ... But variants in this gene cannot be used to assess the effect of quitting ... because people with the risk variants will have smoked more heavily before diagnosis.` The same variant, the same disease, and the instrument is valid for one question and invalid for the next — “Variants in the CHRNA5 gene that associate with increased heaviness of smoking are (through smoking behaviour) associated with increased risk of lung cancer. People with these variants are also likely to find it more difficult to reduce smoking after diagnosis of lung cancer. But variants in this gene cannot be used to assess the effect of quitting or reducing smoking on mortality after diagnosis with lung cancer, because people with the risk variants will have smoked more heavily before diagnosis. The association between CHRNA5 risk variants and mortality in people with lung cancer reflects an individual’s lifetime exposure to tobacco smoke, and we would not expect an instrumental variable estimate using CHRNA5 to reflect the effect of current smoking on prognosis.” | mechanisms[4] |
| mechanism | AND THE DEEPER VERSION OF THAT: `Most importantly, the triggers of disease onset may be entirely different to the factors leading to progression. This means that Mendelian randomisation studies need to be performed using data from studies of disease progression and prognosis, rather than disease occurrence, to generate findings of direct relevance to treating disease.` A causal claim about getting ill is not a causal claim about getting worse, and the same design cannot carry both — “Most importantly, the triggers of disease onset may be entirely different to the factors leading to progression. This means that Mendelian randomisation studies need to be performed using data from studies of disease progression and prognosis, rather than disease occurrence, to generate findings of direct relevance to treating disease.” | mechanisms[5] |
| mechanism | AND A COMMON REPAIR IS CALLED ARBITRARY BY THE GUIDE THAT DESCRIBES IT: `Some studies manually remove (or `PRUNE`) genetic variants thought to be pleiotropic ... Including or excluding a genetic variant in this way can be arbitrary and limited by available data; such pruning should be done with caution (if at all) and for well justified and transparent reasons.` And it can be actively wrong: `Excluding variants on the basis of vertical pleiotropic effects will bias findings, since these effects would be ones seen when manipulating the risk factor of interest.` The two pleiotropies again: prune the wrong one and you delete the signal — “Some studies manually remove (or “prune”) genetic variants thought to be pleiotropic from a genetic instrument to provide a more reliable estimate of the association. Voight and colleagues used genetic variants that were exclusively associated with HDL cholesterol (not LDL cholesterol or triglycerides) to investigate the relationship between HDL cholesterol and heart disease. 60 Including or excluding a genetic variant in this way can be arbitrary and limited by available data; such pruning should be done with caution (if at all) and for well justified and transparent reasons.” | mechanisms[6] |
| mechanism | AND THE TWO ESTIMATES ARE NOT ESTIMATING THE SAME THING, WHICH BREAKS THE OBVIOUS TEST: `Finally, the effects being estimated by the two methods may not be the same - the Mendelian randomisation estimate reflects the effects of LIFELONG PERTURBATIONS in the risk factor, whereas linear regression results may reflect more acute effects. Mendelian randomisation estimates are almost always less precise ..., so tests for difference often have low statistical power.` A disagreement between the two methods can be confounding, or it can be that they answer different questions - and the test that would tell you apart is underpowered — “For continuous outcomes in single sample analyses, the Hausman test can be used to assess whether the Mendelian randomisation and linear regression results (obtained from the same dataset) are systematically different. Differences could occur because the linear regression results are biased by residual confounding, the independence or exclusion restriction assumptions underlying the instrumental variable regression approach are invalid, or both. Finally, the effects being estimated by the two methods may not be the same—the Mendelian randomisation estimate reflects the effects of lifelong perturbations in the risk factor, whereas linear regression results may reflect more acute effects. Mendelian randomisation estimates are almost always less precise and have wider confidence intervals than linear regression ( fig 3 ), so tests for difference often have low statistical power.” | mechanisms[7] |
| mechanism | AND A NULL TEST IS NAMED AS FALSELY REASSURING: `If these tests cannot be rejected, sample size and statistical power may be insufficient to detect differences that do exist and, therefore, may be falsely reassuring.` A heterogeneity test that passes is evidence of nothing, and the guide says so where a reader would otherwise take comfort — “If these tests cannot be rejected, sample size and statistical power may be insufficient to detect differences that do exist and, therefore, may be falsely reassuring.” | mechanisms[8] |
| mechanism | AND THE EVIDENCE HIERARCHY IS USED AND DISOWNED IN ONE SENTENCE: `Although we adapt the conventional pyramid of evidence for presentation purposes, we consider that triangulation of findings from different study designs should be used.` The figure is drawn in the vocabulary the authors say is the wrong one - and the recommendation that follows is a conclusion the pyramid cannot express, because triangulation is not a rank order — “Although we adapt the conventional pyramid of evidence for presentation purposes, we consider that triangulation of findings from different study designs should be used.” | mechanisms[9] |
| mechanism | AND THE LIMIT ON WHAT ANY OF IT LICENSES IS STATED FLATLY: `A well conducted Mendelian randomisation study ... often provides more reliable evidence than a conventional observational study. But the findings should be interpreted in the context of existing evidence from other sources ..., and clinical guidelines should not be rewritten solely on the basis of Mendelian randomisation results.` — “A well conducted Mendelian randomisation study that reasonably satisfies the above assumptions often provides more reliable evidence than a conventional observational study. But the findings should be interpreted in the context of existing evidence from other sources, using different study designs, 75 and clinical guidelines should not be rewritten solely on the basis of Mendelian randomisation results.” | mechanisms[10] |
| extends | the existing guides for conducting and reporting Mendelian randomisation studies, which this paper complements by addressing readers rather than authors — “Mendelian randomisation uses genetic variation as a natural experiment to investigate the causal relations between potentially modifiable risk factors and health outcomes in observational data. As with all epidemiological approaches, findings from Mendelian randomisation studies depend on specific assumptions. We provide explanations of the information typically reported in Mendelian randomisation studies that can be used to assess the plausibility of these assumptions and guidance on how to interpret findings from Mendelian randomisation studies in the context of other sources of evidence” | claim_relations[0] |
| critiques | conventional adjustment for measured confounders, which rests on an assumption the paper calls 'unverifiable' and which measurement error leaves incomplete — “Many epidemiological methods attempt to correct for, or minimise, observed differences in confounders between study participants. These methods can give useful evidence about causal relations if we measure enough confounders so that, after adjustment or matching, study participants who consume different amounts of alcohol are otherwise comparable. But this assumption is unverifiable; if it does not hold, then findings from observational studies will be biasedestimates of causal effects. People who consume more alcohol may also have other risk factors for cardiovascular disease, such as smoking more heavily than those with lower alcohol consumption. The confounding factor (smoking) induces a positive association between the risk factor (alcohol) and an outcome (blood pressure); interpreting this as causal would be misleading. Measuring a confounder does not perfectly characterise it, so measurement error leads to residual confounding, even after apparent statistical adjustment.” | claim_relations[1] |
| critiques | manual pruning of pleiotropic variants, 'arbitrary and limited by available data', and actively biasing when the pleiotropy pruned is vertical — “Some studies manually remove (or “prune”) genetic variants thought to be pleiotropic from a genetic instrument to provide a more reliable estimate of the association. Voight and colleagues used genetic variants that were exclusively associated with HDL cholesterol (not LDL cholesterol or triglycerides) to investigate the relationship between HDL cholesterol and heart disease. 60 Including or excluding a genetic variant in this way can be arbitrary and limited by available data; such pruning should be done with caution (if at all) and for well justified and transparent reasons.” | claim_relations[2] |
| critiques | the conventional pyramid of evidence, adapted 'for presentation purposes' while the authors recommend triangulation instead — “Although we adapt the conventional pyramid of evidence for presentation purposes, we consider that triangulation of findings from different study designs should be used.” | claim_relations[3] |
| sense | `horizontal pleiotropy` - a variant affecting the outcome through a pathway OTHER than the risk factor. `a violation of the instrumental variable assumptions` — “Pleiotropic effects —the effects of a genetic variant on multiple biological pathways. These can either affect the outcome through another trait or pathway to the one under investigation, known as horizontal pleiotropy, or affect other traits through the risk factor of interest, known as vertical pleiotropy. Horizontal pleiotropy is a violation of the instrumental variable assumptions because the effects of the genetic variant on the outcome are not exclusively through the risk factor; this is problematic for Mendelian randomisation studies. Vertical pleiotropy is in fact the essence of Mendelian randomisation—showing that one factor influences a downstream outcome—and is in general not problematic 1” | senses[0] |
| sense | `vertical pleiotropy` - a variant affecting other traits THROUGH the risk factor. `in fact the essence of Mendelian randomisation ... and is in general not problematic` — “Pleiotropic effects —the effects of a genetic variant on multiple biological pathways. These can either affect the outcome through another trait or pathway to the one under investigation, known as horizontal pleiotropy, or affect other traits through the risk factor of interest, known as vertical pleiotropy. Horizontal pleiotropy is a violation of the instrumental variable assumptions because the effects of the genetic variant on the outcome are not exclusively through the risk factor; this is problematic for Mendelian randomisation studies. Vertical pleiotropy is in fact the essence of Mendelian randomisation—showing that one factor influences a downstream outcome—and is in general not problematic 1” | senses[1] |
| sense | `instrumental variable` - `variables that are associated with the risk factor of interest, that are not related to confounders, and that affect the outcome only through the risk factor`; need not be genetic — “Instrumental variables —variables that are associated with the risk factor of interest, that are not related to confounders, and that affect the outcome only through the risk factor. An instrumental variable can be any trait (not necessarily a genetic variant) that meets these criteria, but the nature of genetic inheritance means that genetic variants are often plausible instrumental variables” | senses[2] |
| sense | `weak instrument bias` - what happens with variants `that only explain a small proportion of the variation in the risk factor, coupled with a small sample size` — “Weak instrument bias —can occur in Mendelian randomisation studies when using one or more genetic variants that only explain a small proportion of the variation in the risk factor, coupled with a small sample size” | senses[3] |
| sense | `two sample Mendelian randomisation` - instrument-risk-factor and instrument-outcome associations estimated in DIFFERENT samples; buys power, costs two further assumptions — “Two sample Mendelian randomisation —using two different study samples to estimate the instrument-risk factor and instrument-outcome associations to estimate a causal effect of the risk factor on the outcome. This can be useful when the risk factor or outcome, or both, are expensive to measure. 17 18 It also provides an opportunity to substantially increase the statistical power, by incorporating data from multiple sources, including large consortia 17” | senses[4] |
| sense | `MR Egger regression` - allows pleiotropic effects `as long as the size of these pleiotropic effects is independent of the size of the genetic variants' effects on the risk factor` - a violation permitted under a condition — “MR Egger regression —a statistical technique that allows one or more genetic variants to have pleiotropic effects, as long as the size of these pleiotropic effects is independent of the size of the genetic variants’ effects on the risk factor of interest 19” | senses[5] |
| control | requiring the three instrumental-variable assumptions to hold for EVERY variant, and listing a distinct test for each assumption in single- and two-sample designs — partial — “Valid instrumental variables are defined by three key assumptions ( table 1 , fig 1 ): that they associate with the risk factor of interest (the relevance assumption); that they share no common cause with the outcome (the independence assumption); and that they do not affect the outcome except through the risk factor (the exclusion restriction assumption).” — rules out an instrument that is strong on average while individual variants violate exclusion restriction — “A single genetic variant could plausibly meet these conditions if the biological process linking the variant with the risk factor is well understood. But in many cases Mendelian randomisation studies include multiple genetic variants, which can be used in sensitivity analyses to evaluate the underlying assumptions. Generally, the three key assumptions must hold for each of the genetic variants.” | controls[0] |
| control | MR Egger regression, median and mode estimators, and negative-control populations as sensitivity analyses for pleiotropy — partial — “Furthermore, as the sample size increases, genetic variants will be identified as showing associations with multiple traits owing to vertical or horizontal pleiotropy. Excluding variants on the basis of vertical pleiotropic effects will bias findings, since these effects would be ones seen when manipulating the risk factor of interest.” — rules out horizontal pleiotropy going undetected in a multi-variant instrument — “MR Egger regression —a statistical technique that allows one or more genetic variants to have pleiotropic effects, as long as the size of these pleiotropic effects is independent of the size of the genetic variants’ effects on the risk factor of interest 19” | controls[1] |
| control | warning that a heterogeneity test which cannot be rejected may be underpowered rather than reassuring — held — “In contrast, when the Mendelian randomisation estimate has very wide confidence intervals that overlap the observational estimate and include the null (as in the case of dairy consumption and blood pressure), 69 very little can be inferred from the Mendelian randomisation results.” — rules out reading a null test of difference as evidence of consistency — “If these tests cannot be rejected, sample size and statistical power may be insufficient to detect differences that do exist and, therefore, may be falsely reassuring.” | controls[2] |
| control | stating that clinical guidelines should not be rewritten on Mendelian randomisation alone, and that findings must be read against other designs — held — “Although we adapt the conventional pyramid of evidence for presentation purposes, we consider that triangulation of findings from different study designs should be used.” — rules out a single genetic analysis being taken as decisive for practice — “A well conducted Mendelian randomisation study that reasonably satisfies the above assumptions often provides more reliable evidence than a conventional observational study. But the findings should be interpreted in the context of existing evidence from other sources, using different study designs, 75 and clinical guidelines should not be rewritten solely on the basis of Mendelian randomisation results.” | controls[3] |
| weight | the existing guides for conducting and reporting Mendelian randomisation studies, which this paper complements by addressing readers rather than authors — 0.9 — “Mendelian randomisation uses genetic variation as a natural experiment to investigate the causal relations between potentially modifiable risk factors and health outcomes in observational data. As with all epidemiological approaches, findings from Mendelian randomisation studies depend on specific assumptions. We provide explanations of the information typically reported in Mendelian randomisation studies that can be used to assess the plausibility of these assumptions and guidance on how to interpret findings from Mendelian randomisation studies in the context of other sources of evidence” | claim_relations[0] |
| weight | conventional adjustment for measured confounders, which rests on an assumption the paper calls 'unverifiable' and which measurement error leaves incomplete — 0.9 — “Many epidemiological methods attempt to correct for, or minimise, observed differences in confounders between study participants. These methods can give useful evidence about causal relations if we measure enough confounders so that, after adjustment or matching, study participants who consume different amounts of alcohol are otherwise comparable. But this assumption is unverifiable; if it does not hold, then findings from observational studies will be biasedestimates of causal effects. People who consume more alcohol may also have other risk factors for cardiovascular disease, such as smoking more heavily than those with lower alcohol consumption. The confounding factor (smoking) induces a positive association between the risk factor (alcohol) and an outcome (blood pressure); interpreting this as causal would be misleading. Measuring a confounder does not perfectly characterise it, so measurement error leads to residual confounding, even after apparent statistical adjustment.” | claim_relations[1] |
| weight | manual pruning of pleiotropic variants, 'arbitrary and limited by available data', and actively biasing when the pleiotropy pruned is vertical — 0.85 — “Some studies manually remove (or “prune”) genetic variants thought to be pleiotropic from a genetic instrument to provide a more reliable estimate of the association. Voight and colleagues used genetic variants that were exclusively associated with HDL cholesterol (not LDL cholesterol or triglycerides) to investigate the relationship between HDL cholesterol and heart disease. 60 Including or excluding a genetic variant in this way can be arbitrary and limited by available data; such pruning should be done with caution (if at all) and for well justified and transparent reasons.” | claim_relations[2] |
| weight | the conventional pyramid of evidence, adapted 'for presentation purposes' while the authors recommend triangulation instead — 0.85 — “Although we adapt the conventional pyramid of evidence for presentation purposes, we consider that triangulation of findings from different study designs should be used.” | claim_relations[3] |
| note | horizontal pleiotropy —distinct-from→ vertical pleiotropy — 'These can either affect the outcome through another trait or pathway to the one under investigation, known as horizontal pleiotropy, or affect other traits through the risk factor of interest, known as vertical pleiotropy. Horizontal pleiotropy is a violation of the instrumental variable assumptions because the effects of the genetic variant on the outcome are not exclusively through the risk factor.' The distinction is a direction on an arrow, and it decides whether the paper is evidence or an artefact. Compare 'known as' versus 'also known as' in the corpus: there an adverb split a phrase family; here a modifier splits validity itself — “Pleiotropic effects —the effects of a genetic variant on multiple biological pathways. These can either affect the outcome through another trait or pathway to the one under investigation, known as horizontal pleiotropy, or affect other traits through the risk factor of interest, known as vertical pleiotropy. Horizontal pleiotropy is a violation of the instrumental variable assumptions because the effects of the genetic variant on the outcome are not exclusively through the risk factor; this is problematic for Mendelian randomisation studies. Vertical pleiotropy is in fact the essence of Mendelian randomisation—showing that one factor influences a downstream outcome—and is in general not problematic 1” | concept_edges[0] |
| relation | horizontal pleiotropy —distinct-from→ vertical pleiotropy — both are one variant affecting several traits; horizontal reaches the outcome around the risk factor and breaks the method, vertical reaches it through the risk factor and IS the method — “Pleiotropic effects —the effects of a genetic variant on multiple biological pathways. These can either affect the outcome through another trait or pathway to the one under investigation, known as horizontal pleiotropy, or affect other traits through the risk factor of interest, known as vertical pleiotropy. Horizontal pleiotropy is a violation of the instrumental variable assumptions because the effects of the genetic variant on the outcome are not exclusively through the risk factor; this is problematic for Mendelian randomisation studies. Vertical pleiotropy is in fact the essence of Mendelian randomisation—showing that one factor influences a downstream outcome—and is in general not problematic 1” — “These can either affect the outcome through another trait or pathway to the one under investigation, known as horizontal pleiotropy, or affect other traits through the risk factor of interest, known as vertical pleiotropy. Horizontal pleiotropy is a violation of the instrumental variable assumptions because the effects of the genetic variant on the outcome are not exclusively through the risk factor” | concept_edges[0] |
| weight | horizontal pleiotropy —distinct-from→ vertical pleiotropy — 0.95 — “Pleiotropic effects —the effects of a genetic variant on multiple biological pathways. These can either affect the outcome through another trait or pathway to the one under investigation, known as horizontal pleiotropy, or affect other traits through the risk factor of interest, known as vertical pleiotropy. Horizontal pleiotropy is a violation of the instrumental variable assumptions because the effects of the genetic variant on the outcome are not exclusively through the risk factor; this is problematic for Mendelian randomisation studies. Vertical pleiotropy is in fact the essence of Mendelian randomisation—showing that one factor influences a downstream outcome—and is in general not problematic 1” | concept_edges[0] |
| note | Valid instrumental variables —operationalised-as→ three key assumptions — 'Valid instrumental variables are defined by three key assumptions ...: that they associate with the risk factor of interest (the relevance assumption); that they share no common cause with the outcome (the independence assumption); and that they do not affect the outcome except through the risk factor (the exclusion restriction assumption).' A definition made of three negative or conditional clauses, only the first of which has a statistic attached to it — “Valid instrumental variables are defined by three key assumptions ( table 1 , fig 1 ): that they associate with the risk factor of interest (the relevance assumption); that they share no common cause with the outcome (the independence assumption); and that they do not affect the outcome except through the risk factor (the exclusion restriction assumption).” | concept_edges[1] |
| relation | Valid instrumental variables —operationalised-as→ three key assumptions — validity is exactly relevance, independence and exclusion restriction - and the third is the one that cannot be tested directly — “Valid instrumental variables are defined by three key assumptions ( table 1 , fig 1 ): that they associate with the risk factor of interest (the relevance assumption); that they share no common cause with the outcome (the independence assumption); and that they do not affect the outcome except through the risk factor (the exclusion restriction assumption).” | concept_edges[1] |
| weight | Valid instrumental variables —operationalised-as→ three key assumptions — 0.95 — “Valid instrumental variables are defined by three key assumptions ( table 1 , fig 1 ): that they associate with the risk factor of interest (the relevance assumption); that they share no common cause with the outcome (the independence assumption); and that they do not affect the outcome except through the risk factor (the exclusion restriction assumption).” | concept_edges[1] |
| note | Mendelian randomisation —is-a→ a research method that provides evidence about putative causal relations between modifiable risk factors and disease — 'Mendelian randomisation is a research method that provides evidence about putative causal relations between modifiable risk factors and disease, using genetic variants as natural experiments.' 'Putative' is the hedge the rest of the paper earns: by the end, 'clinical guidelines should not be rewritten solely on the basis of Mendelian randomisation results' — “Mendelian randomisation is a research method that provides evidence about putative causal relations between modifiable risk factors and disease, using genetic variants as natural experiments Mendelian randomisation is less likely to be affected by confounding or reverse causation than conventional observational studies Like all analytical approaches, however, Mendelian randomisation depends on assumptions, and the plausibility of these assumptions must be assessed Moreover, the relevance of the results for clinical decisions should be interpreted in light of other sources of evidence We provide a critical appraisal checklist that can be used to assess and interpret Mendelian randomisation studies” | concept_edges[2] |
| relation | Mendelian randomisation —is-a→ a research method that provides evidence about putative causal relations between modifiable risk factors and disease — the method is presented as a source of evidence about putative causation, using genetic variants as natural experiments - 'putative' doing careful work — “Mendelian randomisation is a research method that provides evidence about putative causal relations between modifiable risk factors and disease, using genetic variants as natural experiments Mendelian randomisation is less likely to be affected by confounding or reverse causation than conventional observational studies Like all analytical approaches, however, Mendelian randomisation depends on assumptions, and the plausibility of these assumptions must be assessed Moreover, the relevance of the results for clinical decisions should be interpreted in light of other sources of evidence We provide a critical appraisal checklist that can be used to assess and interpret Mendelian randomisation studies” — “Mendelian randomisation is a research method that provides evidence about putative causal relations between modifiable risk factors and disease, using genetic variants as natural experiments” | concept_edges[2] |
| weight | Mendelian randomisation —is-a→ a research method that provides evidence about putative causal relations between modifiable risk factors and disease — 0.95 — “Mendelian randomisation is a research method that provides evidence about putative causal relations between modifiable risk factors and disease, using genetic variants as natural experiments Mendelian randomisation is less likely to be affected by confounding or reverse causation than conventional observational studies Like all analytical approaches, however, Mendelian randomisation depends on assumptions, and the plausibility of these assumptions must be assessed Moreover, the relevance of the results for clinical decisions should be interpreted in light of other sources of evidence We provide a critical appraisal checklist that can be used to assess and interpret Mendelian randomisation studies” | concept_edges[2] |
| note | The confounding factor (smoking) —causes→ a positive association between the risk factor (alcohol) and an outcome (blood pressure) — 'The confounding factor (smoking) induces a positive association between the risk factor (alcohol) and an outcome (blood pressure); interpreting this as causal would be misleading.' The clearest textbook statement of confounding in the corpus, and it is the reason the whole method exists — “Many epidemiological methods attempt to correct for, or minimise, observed differences in confounders between study participants. These methods can give useful evidence about causal relations if we measure enough confounders so that, after adjustment or matching, study participants who consume different amounts of alcohol are otherwise comparable. But this assumption is unverifiable; if it does not hold, then findings from observational studies will be biasedestimates of causal effects. People who consume more alcohol may also have other risk factors for cardiovascular disease, such as smoking more heavily than those with lower alcohol consumption. The confounding factor (smoking) induces a positive association between the risk factor (alcohol) and an outcome (blood pressure); interpreting this as causal would be misleading. Measuring a confounder does not perfectly characterise it, so measurement error leads to residual confounding, even after apparent statistical adjustment.” | concept_edges[3] |
| relation | The confounding factor (smoking) —causes→ a positive association between the risk factor (alcohol) and an outcome (blood pressure) — a third variable produces an association between two others, which would be misread as a causal effect of the first on the second — “Many epidemiological methods attempt to correct for, or minimise, observed differences in confounders between study participants. These methods can give useful evidence about causal relations if we measure enough confounders so that, after adjustment or matching, study participants who consume different amounts of alcohol are otherwise comparable. But this assumption is unverifiable; if it does not hold, then findings from observational studies will be biasedestimates of causal effects. People who consume more alcohol may also have other risk factors for cardiovascular disease, such as smoking more heavily than those with lower alcohol consumption. The confounding factor (smoking) induces a positive association between the risk factor (alcohol) and an outcome (blood pressure); interpreting this as causal would be misleading. Measuring a confounder does not perfectly characterise it, so measurement error leads to residual confounding, even after apparent statistical adjustment.” — “The confounding factor (smoking) induces a positive association between the risk factor (alcohol) and an outcome (blood pressure)” | concept_edges[3] |
| weight | The confounding factor (smoking) —causes→ a positive association between the risk factor (alcohol) and an outcome (blood pressure) — 0.95 — “Many epidemiological methods attempt to correct for, or minimise, observed differences in confounders between study participants. These methods can give useful evidence about causal relations if we measure enough confounders so that, after adjustment or matching, study participants who consume different amounts of alcohol are otherwise comparable. But this assumption is unverifiable; if it does not hold, then findings from observational studies will be biasedestimates of causal effects. People who consume more alcohol may also have other risk factors for cardiovascular disease, such as smoking more heavily than those with lower alcohol consumption. The confounding factor (smoking) induces a positive association between the risk factor (alcohol) and an outcome (blood pressure); interpreting this as causal would be misleading. Measuring a confounder does not perfectly characterise it, so measurement error leads to residual confounding, even after apparent statistical adjustment.” | concept_edges[3] |
| note | measurement error —causes→ residual confounding — 'Measuring a confounder does not perfectly characterise it, so measurement error leads to residual confounding, even after apparent statistical adjustment.' This is althubaiti2016's subject arriving as the justification for a different method - measurement error is not only a bias in its own right, it is what makes the standard remedy for another bias incomplete — “Many epidemiological methods attempt to correct for, or minimise, observed differences in confounders between study participants. These methods can give useful evidence about causal relations if we measure enough confounders so that, after adjustment or matching, study participants who consume different amounts of alcohol are otherwise comparable. But this assumption is unverifiable; if it does not hold, then findings from observational studies will be biasedestimates of causal effects. People who consume more alcohol may also have other risk factors for cardiovascular disease, such as smoking more heavily than those with lower alcohol consumption. The confounding factor (smoking) induces a positive association between the risk factor (alcohol) and an outcome (blood pressure); interpreting this as causal would be misleading. Measuring a confounder does not perfectly characterise it, so measurement error leads to residual confounding, even after apparent statistical adjustment.” | concept_edges[4] |
| relation | measurement error —causes→ residual confounding — because a confounder is measured imperfectly, adjusting for the measurement leaves part of the confounding behind — “Many epidemiological methods attempt to correct for, or minimise, observed differences in confounders between study participants. These methods can give useful evidence about causal relations if we measure enough confounders so that, after adjustment or matching, study participants who consume different amounts of alcohol are otherwise comparable. But this assumption is unverifiable; if it does not hold, then findings from observational studies will be biasedestimates of causal effects. People who consume more alcohol may also have other risk factors for cardiovascular disease, such as smoking more heavily than those with lower alcohol consumption. The confounding factor (smoking) induces a positive association between the risk factor (alcohol) and an outcome (blood pressure); interpreting this as causal would be misleading. Measuring a confounder does not perfectly characterise it, so measurement error leads to residual confounding, even after apparent statistical adjustment.” — “Measuring a confounder does not perfectly characterise it, so measurement error leads to residual confounding” | concept_edges[4] |
| weight | measurement error —causes→ residual confounding — 0.95 — “Many epidemiological methods attempt to correct for, or minimise, observed differences in confounders between study participants. These methods can give useful evidence about causal relations if we measure enough confounders so that, after adjustment or matching, study participants who consume different amounts of alcohol are otherwise comparable. But this assumption is unverifiable; if it does not hold, then findings from observational studies will be biasedestimates of causal effects. People who consume more alcohol may also have other risk factors for cardiovascular disease, such as smoking more heavily than those with lower alcohol consumption. The confounding factor (smoking) induces a positive association between the risk factor (alcohol) and an outcome (blood pressure); interpreting this as causal would be misleading. Measuring a confounder does not perfectly characterise it, so measurement error leads to residual confounding, even after apparent statistical adjustment.” | concept_edges[4] |
| note | the Mendelian randomisation estimate —contrasted-with→ linear regression results — 'Finally, the effects being estimated by the two methods may not be the same - the Mendelian randomisation estimate reflects the effects of LIFELONG PERTURBATIONS in the risk factor, whereas linear regression results may reflect more acute effects.' Two numbers that look comparable and are not, and the formal test for their difference 'often ha[s] low statistical power' — “For continuous outcomes in single sample analyses, the Hausman test can be used to assess whether the Mendelian randomisation and linear regression results (obtained from the same dataset) are systematically different. Differences could occur because the linear regression results are biased by residual confounding, the independence or exclusion restriction assumptions underlying the instrumental variable regression approach are invalid, or both. Finally, the effects being estimated by the two methods may not be the same—the Mendelian randomisation estimate reflects the effects of lifelong perturbations in the risk factor, whereas linear regression results may reflect more acute effects. Mendelian randomisation estimates are almost always less precise and have wider confidence intervals than linear regression ( fig 3 ), so tests for difference often have low statistical power.” | concept_edges[5] |
| relation | the Mendelian randomisation estimate —contrasted-with→ linear regression results — the two answer different questions - a lifetime of exposure against an acute effect - so a difference between them is not automatically evidence of confounding — “For continuous outcomes in single sample analyses, the Hausman test can be used to assess whether the Mendelian randomisation and linear regression results (obtained from the same dataset) are systematically different. Differences could occur because the linear regression results are biased by residual confounding, the independence or exclusion restriction assumptions underlying the instrumental variable regression approach are invalid, or both. Finally, the effects being estimated by the two methods may not be the same—the Mendelian randomisation estimate reflects the effects of lifelong perturbations in the risk factor, whereas linear regression results may reflect more acute effects. Mendelian randomisation estimates are almost always less precise and have wider confidence intervals than linear regression ( fig 3 ), so tests for difference often have low statistical power.” — “Finally, the effects being estimated by the two methods may not be the same—the Mendelian randomisation estimate reflects the effects of lifelong perturbations in the risk factor, whereas linear regression results may reflect more acute effects.” | concept_edges[5] |
| weight | the Mendelian randomisation estimate —contrasted-with→ linear regression results — 0.95 — “For continuous outcomes in single sample analyses, the Hausman test can be used to assess whether the Mendelian randomisation and linear regression results (obtained from the same dataset) are systematically different. Differences could occur because the linear regression results are biased by residual confounding, the independence or exclusion restriction assumptions underlying the instrumental variable regression approach are invalid, or both. Finally, the effects being estimated by the two methods may not be the same—the Mendelian randomisation estimate reflects the effects of lifelong perturbations in the risk factor, whereas linear regression results may reflect more acute effects. Mendelian randomisation estimates are almost always less precise and have wider confidence intervals than linear regression ( fig 3 ), so tests for difference often have low statistical power.” | concept_edges[5] |
| note | Variants in the CHRNA5 gene —causes→ increased risk of lung cancer — 'VARIANTS IN THE CHRNA5 GENE that associate with increased heaviness of smoking are (THROUGH SMOKING BEHAVIOUR) associated with increased risk of lung cancer. But variants in this gene cannot be used to assess the effect of quitting ... because people with the risk variants will have smoked more heavily before diagnosis.' Vertical pleiotropy in the first sentence and a broken exclusion restriction in the second, for the same variant - the validity is a property of the question, not of the instrument — “Variants in the CHRNA5 gene that associate with increased heaviness of smoking are (through smoking behaviour) associated with increased risk of lung cancer. People with these variants are also likely to find it more difficult to reduce smoking after diagnosis of lung cancer. But variants in this gene cannot be used to assess the effect of quitting or reducing smoking on mortality after diagnosis with lung cancer, because people with the risk variants will have smoked more heavily before diagnosis. The association between CHRNA5 risk variants and mortality in people with lung cancer reflects an individual’s lifetime exposure to tobacco smoke, and we would not expect an instrumental variable estimate using CHRNA5 to reflect the effect of current smoking on prognosis.” | concept_edges[6] |
| relation | Variants in the CHRNA5 gene —causes→ increased risk of lung cancer — the variants raise heaviness of smoking and, through smoking behaviour, lung cancer risk - a valid instrument for incidence and an invalid one for prognosis — “Variants in the CHRNA5 gene that associate with increased heaviness of smoking are (through smoking behaviour) associated with increased risk of lung cancer. People with these variants are also likely to find it more difficult to reduce smoking after diagnosis of lung cancer. But variants in this gene cannot be used to assess the effect of quitting or reducing smoking on mortality after diagnosis with lung cancer, because people with the risk variants will have smoked more heavily before diagnosis. The association between CHRNA5 risk variants and mortality in people with lung cancer reflects an individual’s lifetime exposure to tobacco smoke, and we would not expect an instrumental variable estimate using CHRNA5 to reflect the effect of current smoking on prognosis.” — “Variants in the CHRNA5 gene that associate with increased heaviness of smoking are (through smoking behaviour) associated with increased risk of lung cancer.” | concept_edges[6] |
| weight | Variants in the CHRNA5 gene —causes→ increased risk of lung cancer — 0.9 — “Variants in the CHRNA5 gene that associate with increased heaviness of smoking are (through smoking behaviour) associated with increased risk of lung cancer. People with these variants are also likely to find it more difficult to reduce smoking after diagnosis of lung cancer. But variants in this gene cannot be used to assess the effect of quitting or reducing smoking on mortality after diagnosis with lung cancer, because people with the risk variants will have smoked more heavily before diagnosis. The association between CHRNA5 risk variants and mortality in people with lung cancer reflects an individual’s lifetime exposure to tobacco smoke, and we would not expect an instrumental variable estimate using CHRNA5 to reflect the effect of current smoking on prognosis.” | concept_edges[6] |
| audit | the triggers of disease onset —distinct-from→ the factors leading to progression — Audit 2026-09-28: DOUBTFUL - read against its full span and kept, but the triggers of onset 'may be entirely different' from those of progression; hedged. — “Most importantly, the triggers of disease onset may be entirely different to the factors leading to progression. This means that Mendelian randomisation studies need to be performed using data from studies of disease progression and prognosis, rather than disease occurrence, to generate findings of direct relevance to treating disease.” | concept_edges[7] |
| note | the triggers of disease onset —distinct-from→ the factors leading to progression — 'Most importantly, the triggers of disease onset may be entirely different to the factors leading to progression. This means that Mendelian randomisation studies need to be performed using data from studies of disease progression and prognosis, rather than disease occurrence.' Nearly all the available data is occurrence data, so this is a limitation on the evidence base rather than on the method — “Most importantly, the triggers of disease onset may be entirely different to the factors leading to progression. This means that Mendelian randomisation studies need to be performed using data from studies of disease progression and prognosis, rather than disease occurrence, to generate findings of direct relevance to treating disease.” | concept_edges[7] |
| relation | the triggers of disease onset —distinct-from→ the factors leading to progression — what causes a disease to start may be unrelated to what makes it progress, so studies of occurrence cannot answer questions about treatment — “Most importantly, the triggers of disease onset may be entirely different to the factors leading to progression. This means that Mendelian randomisation studies need to be performed using data from studies of disease progression and prognosis, rather than disease occurrence, to generate findings of direct relevance to treating disease.” — “Most importantly, the triggers of disease onset may be entirely different to the factors leading to progression.” | concept_edges[7] |
| weight | the triggers of disease onset —distinct-from→ the factors leading to progression — 0.95 — “Most importantly, the triggers of disease onset may be entirely different to the factors leading to progression. This means that Mendelian randomisation studies need to be performed using data from studies of disease progression and prognosis, rather than disease occurrence, to generate findings of direct relevance to treating disease.” | concept_edges[7] |
| outcome | control — partial — “A single genetic variant could plausibly meet these conditions if the biological process linking the variant with the risk factor is well understood. But in many cases Mendelian randomisation studies include multiple genetic variants, which can be used in sensitivity analyses to evaluate the underlying assumptions. Generally, the three key assumptions must hold for each of the genetic variants.” — “Valid instrumental variables are defined by three key assumptions ( table 1 , fig 1 ): that they associate with the risk factor of interest (the relevance assumption); that they share no common cause with the outcome (the independence assumption); and that they do not affect the outcome except through the risk factor (the exclusion restriction assumption).” | controls[0] |
| outcome | control — partial — “MR Egger regression —a statistical technique that allows one or more genetic variants to have pleiotropic effects, as long as the size of these pleiotropic effects is independent of the size of the genetic variants’ effects on the risk factor of interest 19” — “Furthermore, as the sample size increases, genetic variants will be identified as showing associations with multiple traits owing to vertical or horizontal pleiotropy. Excluding variants on the basis of vertical pleiotropic effects will bias findings, since these effects would be ones seen when manipulating the risk factor of interest.” | controls[1] |
| outcome | control — held — “If these tests cannot be rejected, sample size and statistical power may be insufficient to detect differences that do exist and, therefore, may be falsely reassuring.” — “In contrast, when the Mendelian randomisation estimate has very wide confidence intervals that overlap the observational estimate and include the null (as in the case of dairy consumption and blood pressure), 69 very little can be inferred from the Mendelian randomisation results.” | controls[2] |
| outcome | control — held — “A well conducted Mendelian randomisation study that reasonably satisfies the above assumptions often provides more reliable evidence than a conventional observational study. But the findings should be interpreted in the context of existing evidence from other sources, using different study designs, 75 and clinical guidelines should not be rewritten solely on the basis of Mendelian randomisation results.” — “Although we adapt the conventional pyramid of evidence for presentation purposes, we consider that triangulation of findings from different study designs should be used.” | controls[3] |
| limitation | EVERY ASSUMPTION IS PER-VARIANT AND MOST STUDIES USE MANY VARIANTS: `Generally, the three key assumptions must hold for each of the genetic variants.` A study with a hundred-variant allele score is making three hundred assumptions, of which the exclusion restriction is `NOT DIRECTLY TESTABLE` in the general case - the tools listed against it are biological knowledge, association tests against alternative pathways, and MR Egger — “A single genetic variant could plausibly meet these conditions if the biological process linking the variant with the risk factor is well understood. But in many cases Mendelian randomisation studies include multiple genetic variants, which can be used in sensitivity analyses to evaluate the underlying assumptions. Generally, the three key assumptions must hold for each of the genetic variants.” | limitations[0] |
| limitation | AND THE FIX FOR PLEIOTROPY IS ITSELF ARBITRARY: `Including or excluding a genetic variant in this way can be arbitrary and limited by available data; such pruning should be done with caution (if at all)`, and `Excluding variants on the basis of vertical pleiotropic effects will bias findings`. The remedy and the damage differ by which pleiotropy you think you are looking at, and that judgement is not testable from the data — “Furthermore, as the sample size increases, genetic variants will be identified as showing associations with multiple traits owing to vertical or horizontal pleiotropy. Excluding variants on the basis of vertical pleiotropic effects will bias findings, since these effects would be ones seen when manipulating the risk factor of interest.” | limitations[1] |
| limitation | AND THE ESTIMATE IS LIFELONG WHILE THE DECISION IS NOT: `the numerical effect estimates may not be clinically meaningful. Will interventions at a specific age have the same sized effects?` The vitamin D and multiple sclerosis case is given as the example - `such protection may be limited to childhood and adolescence and not later adult life` — “But this must be approached with care because genetic variants generally relate to lifelong differences in a risk factor, not the effects at a specific time. ALDH2 is associated with alcohol consumption from adolescence through middle age, 76 so the Mendelian randomisation results (coupled with biological knowledge) indicate that consuming less alcohol would lead to lower blood pressure on average, but do not necessarily inform the effect size of an intervention at a specific time in life for a specific duration.” | limitations[2] |
| limitation | AND THE TWO-SAMPLE DESIGN BUYS POWER AND ADDS ASSUMPTIONS: `These advantages come with two additional assumptions: the two samples are assumed to represent the same underlying population, and overlap in participants between the two samples can cause bias towards the risk factor-outcome association.` A design chosen because the data is cheaper requires two claims about populations nobody measured together — “These advantages come with two additional assumptions: the two samples are assumed to represent the same underlying population, and overlap in participants between the two samples can cause bias towards the risk factor-outcome association.” | limitations[3] |
| limitation | AND WIDE INTERVALS ARE NOT A NULL: `when the Mendelian randomisation estimate has very wide confidence intervals that overlap the observational estimate and include the null ..., very little can be inferred from the Mendelian randomisation results.` Read with the falsely-reassuring heterogeneity test, both of the paper's warnings are about mistaking imprecision for agreement — “In contrast, when the Mendelian randomisation estimate has very wide confidence intervals that overlap the observational estimate and include the null (as in the case of dairy consumption and blood pressure), 69 very little can be inferred from the Mendelian randomisation results.” | limitations[4] |
| limitation | AND THE METHODOLOGICAL SETTLEMENT IS OPEN: `The conversation on how to conduct, report, and interpret Mendelian randomisation studies is still ongoing.` A guide with 4,781 citations, saying that the thing it is a guide to is not settled — “The conversation on how to conduct, report, and interpret Mendelian randomisation studies is still ongoing. Other areas of medical research, such as randomised controlled trials, have been greatly strengthened by close collaborations between methodologists, empirical researchers, and clinicians. Similar collaborations are needed to ensure the strengths and limitations of Mendelian randomisation are fully appreciated and realised.” | limitations[5] |
finding · quantity · variable · definition · key-term — grounded in the full text| finding | — “Pleiotropic effects —the effects of a genetic variant on multiple biological pathways. These can either affect the outcome through another trait or pathway to the one under investigation, known as horizontal pleiotropy, or affect other traits through the risk factor of interest, known as vertical pleiotropy. Horizontal pleiotropy is a violation of the instrumental variable assumptions because the effects of the genetic variant on the outcome are not exclusively through the risk factor; this is problematic for Mendelian randomisation studies. Vertical pleiotropy is in fact the essence of Mendelian randomisation—showing that one factor influences a downstream outcome—and is in general not problematic 1” | content_tags.findings[0] |
| finding | — “Valid instrumental variables are defined by three key assumptions ( table 1 , fig 1 ): that they associate with the risk factor of interest (the relevance assumption); that they share no common cause with the outcome (the independence assumption); and that they do not affect the outcome except through the risk factor (the exclusion restriction assumption).” | content_tags.findings[1] |
| finding | — “But this must be approached with care because genetic variants generally relate to lifelong differences in a risk factor, not the effects at a specific time. ALDH2 is associated with alcohol consumption from adolescence through middle age, 76 so the Mendelian randomisation results (coupled with biological knowledge) indicate that consuming less alcohol would lead to lower blood pressure on average, but do not necessarily inform the effect size of an intervention at a specific time in life for a specific duration.” | content_tags.findings[2] |
| finding | — “Variants in the CHRNA5 gene that associate with increased heaviness of smoking are (through smoking behaviour) associated with increased risk of lung cancer. People with these variants are also likely to find it more difficult to reduce smoking after diagnosis of lung cancer. But variants in this gene cannot be used to assess the effect of quitting or reducing smoking on mortality after diagnosis with lung cancer, because people with the risk variants will have smoked more heavily before diagnosis. The association between CHRNA5 risk variants and mortality in people with lung cancer reflects an individual’s lifetime exposure to tobacco smoke, and we would not expect an instrumental variable estimate using CHRNA5 to reflect the effect of current smoking on prognosis.” | content_tags.findings[3] |
| finding | — “Most importantly, the triggers of disease onset may be entirely different to the factors leading to progression. This means that Mendelian randomisation studies need to be performed using data from studies of disease progression and prognosis, rather than disease occurrence, to generate findings of direct relevance to treating disease.” | content_tags.findings[4] |
| finding | — “Some studies manually remove (or “prune”) genetic variants thought to be pleiotropic from a genetic instrument to provide a more reliable estimate of the association. Voight and colleagues used genetic variants that were exclusively associated with HDL cholesterol (not LDL cholesterol or triglycerides) to investigate the relationship between HDL cholesterol and heart disease. 60 Including or excluding a genetic variant in this way can be arbitrary and limited by available data; such pruning should be done with caution (if at all) and for well justified and transparent reasons.” | content_tags.findings[5] |
| finding | — “If these tests cannot be rejected, sample size and statistical power may be insufficient to detect differences that do exist and, therefore, may be falsely reassuring.” | content_tags.findings[6] |
| finding | — “Although we adapt the conventional pyramid of evidence for presentation purposes, we consider that triangulation of findings from different study designs should be used.” | content_tags.findings[7] |
| finding | — “The conversation on how to conduct, report, and interpret Mendelian randomisation studies is still ongoing. Other areas of medical research, such as randomised controlled trials, have been greatly strengthened by close collaborations between methodologists, empirical researchers, and clinicians. Similar collaborations are needed to ensure the strengths and limitations of Mendelian randomisation are fully appreciated and realised.” | content_tags.findings[8] |
| quantity | 3 per genetic variant - a hundred-variant instrument makes three hundred — — “A single genetic variant could plausibly meet these conditions if the biological process linking the variant with the risk factor is well understood. But in many cases Mendelian randomisation studies include multiple genetic variants, which can be used in sensitivity analyses to evaluate the underlying assumptions. Generally, the three key assumptions must hold for each of the genetic variants.” | content_tags.quantities[0] |
| quantity | ALDH2: 170 of 1919 men with two A alleles drank 1.1 g of alcohol a day against 23.7 g for those with none — — “In the case of alcohol and blood pressure, a variant in the ALDH2 gene (specifically the minor A allele of rs671, rather than the wild type or major allele G) found in east Asian populations slows the metabolism of acetaldehyde, which causes a flush response and other adverse responses to alcohol consumption. In a study of 4057 people selected from the general population, 170 of 1919 men carried two copies of the A allele and drank an average of 1.1 g of alcohol a day, whereas those with no copies drank 23.7 g.” | content_tags.quantities[1] |
| quantity | 2 - same underlying population, and no participant overlap — — “These advantages come with two additional assumptions: the two samples are assumed to represent the same underlying population, and overlap in participants between the two samples can cause bias towards the risk factor-outcome association.” | content_tags.quantities[2] |
| variable | the genetic instrument — “Instrumental variables —variables that are associated with the risk factor of interest, that are not related to confounders, and that affect the outcome only through the risk factor. An instrumental variable can be any trait (not necessarily a genetic variant) that meets these criteria, but the nature of genetic inheritance means that genetic variants are often plausible instrumental variables” | content_tags.variables[0] |
| variable | overlap in participants — “These advantages come with two additional assumptions: the two samples are assumed to represent the same underlying population, and overlap in participants between the two samples can cause bias towards the risk factor-outcome association.” | content_tags.variables[1] |
| definition | instrumental variable — “Instrumental variables —variables that are associated with the risk factor of interest, that are not related to confounders, and that affect the outcome only through the risk factor. An instrumental variable can be any trait (not necessarily a genetic variant) that meets these criteria, but the nature of genetic inheritance means that genetic variants are often plausible instrumental variables” | content_tags.definitions[0] |
| definition | the three key assumptions — “Valid instrumental variables are defined by three key assumptions ( table 1 , fig 1 ): that they associate with the risk factor of interest (the relevance assumption); that they share no common cause with the outcome (the independence assumption); and that they do not affect the outcome except through the risk factor (the exclusion restriction assumption).” | content_tags.definitions[2] |
| definition | weak instrument bias — “Weak instrument bias —can occur in Mendelian randomisation studies when using one or more genetic variants that only explain a small proportion of the variation in the risk factor, coupled with a small sample size” | content_tags.definitions[3] |
| definition | two sample Mendelian randomisation — “Two sample Mendelian randomisation —using two different study samples to estimate the instrument-risk factor and instrument-outcome associations to estimate a causal effect of the risk factor on the outcome. This can be useful when the risk factor or outcome, or both, are expensive to measure. 17 18 It also provides an opportunity to substantially increase the statistical power, by incorporating data from multiple sources, including large consortia 17” | content_tags.definitions[4] |
| definition | horizontal pleiotropy — “Pleiotropic effects —the effects of a genetic variant on multiple biological pathways. These can either affect the outcome through another trait or pathway to the one under investigation, known as horizontal pleiotropy, or affect other traits through the risk factor of interest, known as vertical pleiotropy. Horizontal pleiotropy is a violation of the instrumental variable assumptions because the effects of the genetic variant on the outcome are not exclusively through the risk factor; this is problematic for Mendelian randomisation studies. Vertical pleiotropy is in fact the essence of Mendelian randomisation—showing that one factor influences a downstream outcome—and is in general not problematic 1” | content_tags.definitions[5] |
| definition | vertical pleiotropy — “Pleiotropic effects —the effects of a genetic variant on multiple biological pathways. These can either affect the outcome through another trait or pathway to the one under investigation, known as horizontal pleiotropy, or affect other traits through the risk factor of interest, known as vertical pleiotropy. Horizontal pleiotropy is a violation of the instrumental variable assumptions because the effects of the genetic variant on the outcome are not exclusively through the risk factor; this is problematic for Mendelian randomisation studies. Vertical pleiotropy is in fact the essence of Mendelian randomisation—showing that one factor influences a downstream outcome—and is in general not problematic 1” | content_tags.definitions[6] |
| concept | horizontal pleiotropy — not defined in this paper's own words; the effect that invalidates the instrument — “Pleiotropic effects —the effects of a genetic variant on multiple biological pathways. These can either affect the outcome through another trait or pathway to the one under investigation, known as horizontal pleiotropy, or affect other traits through the risk factor of interest, known as vertical pleiotropy. Horizontal pleiotropy is a violation of the instrumental variable assumptions because the effects of the genetic variant on the outcome are not exclusively through the risk factor; this is problematic for Mendelian randomisation studies. Vertical pleiotropy is in fact the essence of Mendelian randomisation—showing that one factor influences a downstream outcome—and is in general not problematic 1” | concepts[0] |
| concept | vertical pleiotropy — not defined in this paper's own words; the effect that is the method working — “Pleiotropic effects —the effects of a genetic variant on multiple biological pathways. These can either affect the outcome through another trait or pathway to the one under investigation, known as horizontal pleiotropy, or affect other traits through the risk factor of interest, known as vertical pleiotropy. Horizontal pleiotropy is a violation of the instrumental variable assumptions because the effects of the genetic variant on the outcome are not exclusively through the risk factor; this is problematic for Mendelian randomisation studies. Vertical pleiotropy is in fact the essence of Mendelian randomisation—showing that one factor influences a downstream outcome—and is in general not problematic 1” | concepts[1] |
| concept | Valid instrumental variables — not defined in this paper's own words; the objects the method requires — “Valid instrumental variables are defined by three key assumptions ( table 1 , fig 1 ): that they associate with the risk factor of interest (the relevance assumption); that they share no common cause with the outcome (the independence assumption); and that they do not affect the outcome except through the risk factor (the exclusion restriction assumption).” | concepts[2] |
| concept | three key assumptions — not defined in this paper's own words; relevance, independence, exclusion restriction — “Valid instrumental variables are defined by three key assumptions ( table 1 , fig 1 ): that they associate with the risk factor of interest (the relevance assumption); that they share no common cause with the outcome (the independence assumption); and that they do not affect the outcome except through the risk factor (the exclusion restriction assumption).” | concepts[3] |
| concept | Mendelian randomisation — not defined in this paper's own words; the method under appraisal — “Mendelian randomisation is a research method that provides evidence about putative causal relations between modifiable risk factors and disease, using genetic variants as natural experiments Mendelian randomisation is less likely to be affected by confounding or reverse causation than conventional observational studies Like all analytical approaches, however, Mendelian randomisation depends on assumptions, and the plausibility of these assumptions must be assessed Moreover, the relevance of the results for clinical decisions should be interpreted in light of other sources of evidence We provide a critical appraisal checklist that can be used to assess and interpret Mendelian randomisation studies” | concepts[4] |
| concept | a research method that provides evidence about putative causal relations between modifiable risk factors and disease — not defined in this paper's own words; what it is said to be — “Mendelian randomisation is a research method that provides evidence about putative causal relations between modifiable risk factors and disease, using genetic variants as natural experiments Mendelian randomisation is less likely to be affected by confounding or reverse causation than conventional observational studies Like all analytical approaches, however, Mendelian randomisation depends on assumptions, and the plausibility of these assumptions must be assessed Moreover, the relevance of the results for clinical decisions should be interpreted in light of other sources of evidence We provide a critical appraisal checklist that can be used to assess and interpret Mendelian randomisation studies” | concepts[5] |
| concept | The confounding factor (smoking) — not defined in this paper's own words; the worked example of confounding — “Many epidemiological methods attempt to correct for, or minimise, observed differences in confounders between study participants. These methods can give useful evidence about causal relations if we measure enough confounders so that, after adjustment or matching, study participants who consume different amounts of alcohol are otherwise comparable. But this assumption is unverifiable; if it does not hold, then findings from observational studies will be biasedestimates of causal effects. People who consume more alcohol may also have other risk factors for cardiovascular disease, such as smoking more heavily than those with lower alcohol consumption. The confounding factor (smoking) induces a positive association between the risk factor (alcohol) and an outcome (blood pressure); interpreting this as causal would be misleading. Measuring a confounder does not perfectly characterise it, so measurement error leads to residual confounding, even after apparent statistical adjustment.” | concepts[6] |
| concept | a positive association between the risk factor (alcohol) and an outcome (blood pressure) — not defined in this paper's own words; what it induces, and what would be misread as causal — “Many epidemiological methods attempt to correct for, or minimise, observed differences in confounders between study participants. These methods can give useful evidence about causal relations if we measure enough confounders so that, after adjustment or matching, study participants who consume different amounts of alcohol are otherwise comparable. But this assumption is unverifiable; if it does not hold, then findings from observational studies will be biasedestimates of causal effects. People who consume more alcohol may also have other risk factors for cardiovascular disease, such as smoking more heavily than those with lower alcohol consumption. The confounding factor (smoking) induces a positive association between the risk factor (alcohol) and an outcome (blood pressure); interpreting this as causal would be misleading. Measuring a confounder does not perfectly characterise it, so measurement error leads to residual confounding, even after apparent statistical adjustment.” | concepts[7] |
| concept | measurement error — not defined in this paper's own words; why adjustment cannot finish the job — “Many epidemiological methods attempt to correct for, or minimise, observed differences in confounders between study participants. These methods can give useful evidence about causal relations if we measure enough confounders so that, after adjustment or matching, study participants who consume different amounts of alcohol are otherwise comparable. But this assumption is unverifiable; if it does not hold, then findings from observational studies will be biasedestimates of causal effects. People who consume more alcohol may also have other risk factors for cardiovascular disease, such as smoking more heavily than those with lower alcohol consumption. The confounding factor (smoking) induces a positive association between the risk factor (alcohol) and an outcome (blood pressure); interpreting this as causal would be misleading. Measuring a confounder does not perfectly characterise it, so measurement error leads to residual confounding, even after apparent statistical adjustment.” | concepts[8] |
| concept | residual confounding — not defined in this paper's own words; what is left after apparent statistical adjustment — “Many epidemiological methods attempt to correct for, or minimise, observed differences in confounders between study participants. These methods can give useful evidence about causal relations if we measure enough confounders so that, after adjustment or matching, study participants who consume different amounts of alcohol are otherwise comparable. But this assumption is unverifiable; if it does not hold, then findings from observational studies will be biasedestimates of causal effects. People who consume more alcohol may also have other risk factors for cardiovascular disease, such as smoking more heavily than those with lower alcohol consumption. The confounding factor (smoking) induces a positive association between the risk factor (alcohol) and an outcome (blood pressure); interpreting this as causal would be misleading. Measuring a confounder does not perfectly characterise it, so measurement error leads to residual confounding, even after apparent statistical adjustment.” | concepts[9] |
| concept | the Mendelian randomisation estimate — not defined in this paper's own words; an effect of lifelong perturbation — “For continuous outcomes in single sample analyses, the Hausman test can be used to assess whether the Mendelian randomisation and linear regression results (obtained from the same dataset) are systematically different. Differences could occur because the linear regression results are biased by residual confounding, the independence or exclusion restriction assumptions underlying the instrumental variable regression approach are invalid, or both. Finally, the effects being estimated by the two methods may not be the same—the Mendelian randomisation estimate reflects the effects of lifelong perturbations in the risk factor, whereas linear regression results may reflect more acute effects. Mendelian randomisation estimates are almost always less precise and have wider confidence intervals than linear regression ( fig 3 ), so tests for difference often have low statistical power.” | concepts[10] |
| concept | linear regression results — not defined in this paper's own words; an effect that may be acute — “For continuous outcomes in single sample analyses, the Hausman test can be used to assess whether the Mendelian randomisation and linear regression results (obtained from the same dataset) are systematically different. Differences could occur because the linear regression results are biased by residual confounding, the independence or exclusion restriction assumptions underlying the instrumental variable regression approach are invalid, or both. Finally, the effects being estimated by the two methods may not be the same—the Mendelian randomisation estimate reflects the effects of lifelong perturbations in the risk factor, whereas linear regression results may reflect more acute effects. Mendelian randomisation estimates are almost always less precise and have wider confidence intervals than linear regression ( fig 3 ), so tests for difference often have low statistical power.” | concepts[11] |
| concept | Variants in the CHRNA5 gene — not defined in this paper's own words; the instrument that answers one question and not the next — “Variants in the CHRNA5 gene that associate with increased heaviness of smoking are (through smoking behaviour) associated with increased risk of lung cancer. People with these variants are also likely to find it more difficult to reduce smoking after diagnosis of lung cancer. But variants in this gene cannot be used to assess the effect of quitting or reducing smoking on mortality after diagnosis with lung cancer, because people with the risk variants will have smoked more heavily before diagnosis. The association between CHRNA5 risk variants and mortality in people with lung cancer reflects an individual’s lifetime exposure to tobacco smoke, and we would not expect an instrumental variable estimate using CHRNA5 to reflect the effect of current smoking on prognosis.” | concepts[12] |
| concept | increased risk of lung cancer — not defined in this paper's own words; what they are associated with, through smoking behaviour — “Variants in the CHRNA5 gene that associate with increased heaviness of smoking are (through smoking behaviour) associated with increased risk of lung cancer. People with these variants are also likely to find it more difficult to reduce smoking after diagnosis of lung cancer. But variants in this gene cannot be used to assess the effect of quitting or reducing smoking on mortality after diagnosis with lung cancer, because people with the risk variants will have smoked more heavily before diagnosis. The association between CHRNA5 risk variants and mortality in people with lung cancer reflects an individual’s lifetime exposure to tobacco smoke, and we would not expect an instrumental variable estimate using CHRNA5 to reflect the effect of current smoking on prognosis.” | concepts[13] |
| concept | the triggers of disease onset — not defined in this paper's own words; what most Mendelian randomisation studies are powered on — “Most importantly, the triggers of disease onset may be entirely different to the factors leading to progression. This means that Mendelian randomisation studies need to be performed using data from studies of disease progression and prognosis, rather than disease occurrence, to generate findings of direct relevance to treating disease.” | concepts[14] |
| concept | the factors leading to progression — not defined in this paper's own words; what a clinician treating a patient needs, and a different question — “Most importantly, the triggers of disease onset may be entirely different to the factors leading to progression. This means that Mendelian randomisation studies need to be performed using data from studies of disease progression and prognosis, rather than disease occurrence, to generate findings of direct relevance to treating disease.” | concepts[15] |
| a test for difference that is underpowered by construction | — “For continuous outcomes in single sample analyses, the Hausman test can be used to assess whether the Mendelian randomisation and linear regression results (obtained from the same dataset) are systematically different. Differences could occur because the linear regression results are biased by residual confounding, the independence or exclusion restriction assumptions underlying the instrumental variable regression approach are invalid, or both. Finally, the effects being estimated by the two methods may not be the same—the Mendelian randomisation estimate reflects the effects of lifelong perturbations in the risk factor, whereas linear regression results may reflect more acute effects. Mendelian randomisation estimates are almost always less precise and have wider confidence intervals than linear regression ( fig 3 ), so tests for difference often have low statistical power.” | content_tags.statistics[0] |
| wide intervals are not a null result | — “In contrast, when the Mendelian randomisation estimate has very wide confidence intervals that overlap the observational estimate and include the null (as in the case of dairy consumption and blood pressure), 69 very little can be inferred from the Mendelian randomisation results.” | content_tags.statistics[1] |
Full text — tagged
The reference list is not part of the text we hold for this paper. Our copy was extracted from the publisher's page and the bibliography did not survive that extraction, so the article below ends before its references. Nothing has been removed by us.
BMJ BMJ 3 bmj BMJ-UK The BMJ 0959-8138 1756-1833 BMJ Publishing Group PMC6041728 PMC6041728.1 6041728 6041728 30002074 10.1136/bmj.k601 davn040667 1 Research Methods & Reporting Reading Mendelian randomisationterm studies: a guide, glossary, and checklist for clinicians DaviespersName Neil M research fellow 1 2 HolmespersName Michael V university research lecturer 1 3 4 5 Davey SmithpersName George professor 1 2 6 1 Medical Research Council Integrative Epidemiology Unit, University of Bristol, BS8 2BN, UK 2 Population Health Sciences, Bristol Medical School, University of Bristol, Barley House, Oakfield Grove, Bristol, BS8 2BN, UK 3 Medical Research Council Population Health Research Unit, University of Oxford, UK 4 Clinical Trial Service Unit and Epidemiological Studies Unit (CTSU), Nuffield Department of Population Health, University of Oxford, Richard Doll Building, Old Road Campus, Roosevelt Drive, Oxford OX3 7LF, UK. National Institute for Health Research Oxford Biomedical Research Centre, Oxford University Hospital, Oxford, UK 6 National Institute for Health Research Bristol Biomedical Research Centre, Oakfield House, Oakfield Grove, Bristol BS8 2BN, UK Correspondence: G Davey SmithpersName pa-ieudirector@bristol.ac.uk 2018 12 7 2018 362 316247 k601 04 12 2017 01 01 2018 13 07 2018 07 07 2025 Published by the BMJ Publishing Group Limited. For permission to use (where not already granted under a licence) please go to http://group.bmj.com/group/rights-licensing/permissions 2018 BMJ https://creativecommons.org/licenses/by/4.0/ This is an Open Access article distributed in accordance with the terms of the Creative Commons Attribution (CC BY 4.0) license, which permits others to distribute, remix, adapt and build upon this work, for commercial use, provided the original work is properly cited. See: http://creativecommons.org/licenses/by/4.0/ . Mendelian randomisationterm uses genetic variation as a natural experiment to investigate the causal relations between potentially modifiable risk factors and health outcomes in observational data. As with all epidemiological approaches, findings from Mendelian randomisationterm studies depend on specific assumptions. We provide explanations of the information typically reported in Mendelian randomisationterm studies that can be used to assess the plausibility of these assumptions and guidance on how to interpret findings from Mendelian randomisationterm studies in the context of other sources of evidence pmc-status-qastatus 0 pmc-status-live yes pmc-status-embargo no pmc-status-released yes pmc-prop-open-access yes pmc-prop-olf no pmc-prop-manuscript no pmc-prop-legally-suppressed no pmc-prop-has-pdf no pmc-prop-has-supplement yes pmc-prop-pdf-only no pmc-prop-suppress-copyright no pmc-prop-is-real-version no pmc-prop-is-scanned-article no pmc-prop-preprint no pmc-prop-in-epmc yes pmc-license-ref CC BY Summary points Mendelian randomisationterm is a research method that provides evidence about putative causal relations between modifiable risk factors and diseaseterm, using genetic variants as natural experiments Mendelian randomisationterm is less likely to be affected by confounding or reverse causation than conventional observational studies Like all analytical approaches, however, Mendelian randomisationterm depends on assumptions, and the plausibility of these assumptions must be assessed Moreover, the relevance of the results for clinical decisions should be interpreted in light of other sources of evidence We provide a critical appraisal checklist that can be used to assess and interpret Mendelian randomisationterm studies Understanding whether a biomarker or behaviour causes ill health is central to evidence based medicine, drug development, and better informed clinical decision making. Ideally, evidence of causal effects comes from well conducted randomised trials. Clinicians are well versed in the strengths and limitations of such trials and have an increasingly sophisticated understanding of traditional analyses of observational studies. But they may be less aware of the strengths and limitations of a more recently developed approach to analysing observational data known as Mendelian randomisationterm. Although numerous guides exist for conducting 1 2 3 and reporting Mendelian randomisationterm studies and related methods, 4 5 here we focus on helping clinicians and practitioners read and interpret them. Our goal is to provide explanations of core concepts and recent developments in Mendelian randomisationterm methods.
A method to overcome confounding Mendelian randomisationterm is an analytical method that uses genetic variants as instrumental variablesterm·def for modifiable risk factors that affect population health. 1 6 7 8 It is increasingly being used because it can overcome a major limitation of evidence from observational studies: unmeasured confounding. 6 9 Suppose, for example, we wanted to investigate the effects of alcohol consumption on blood pressure with a view to understanding the overall relationship of alcohol with risk of coronary heart disease. One source of evidence is the association between alcohol and blood pressure in observational studies. This association may be a poor indicator of the causal effects of alcohol if there are other factors—“confounders”—that influence both alcohol intake and blood pressure. Many epidemiological methods attempt to correct for, or minimise, observed differences in confounders between study participants. These methods can give useful evidence about causal relations if we measure enough confounders so that, after adjustment or matching, study participants who consume different amounts of alcohol are otherwise comparable. But this assumption is unverifiable; if it does not hold, then findings from observational studies will be biasedestimates of causal effects.
People who consume more alcohol may also have other risk factors for cardiovascular disease, such as smoking more heavily than those with lower alcohol consumption. The confounding factor (smoking)term induces a positive association between the risk factor (alcohol) and an outcome (blood pressure)term; interpreting this as causal would be misleading. Measuring a confounder does not perfectly characterise it, so measurement errorterm leads to residual confoundingterm, even after apparent statistical adjustment. Reverse causality is a form of confounding that is difficult to account for. It arises if the outcome or preclinical aspects of the disease that lead to the outcome affect the risk factor. People with symptoms of cardiovascular disease, for example, may consume less alcohol than those without symptoms. This would lead to a negative association between a risk factor (alcohol) and an outcome (cardiovascular disease); interpreting this as being because alcohol consumption decreases the risk of cardiovascular disease would be misleading.
Mendelian randomisationterm uses genetic variants, which are fixed at conception, to support causal inferences about the effects of modifiable risk factors, which can overcome some types of confounding. In the case of alcohol and blood pressure, a variant in the ALDH2 gene (specifically the minor A allele of rs671, rather than the wild type or major allele G) found in east Asian populations slows the metabolism of acetaldehyde, which causes a flush response and other adverse responses to alcohol consumption. In a study of 4057 people selected from the general population, 170 of 1919 men carried two copies of the A allele and drank an average of 1.1 g of alcohol a day, whereas those with no copies drank 23.7 g. 10 11 If men with one or more copies of the A allele have lower blood pressure, then this implies that lower alcohol consumption decreases blood pressure. 11 But this inference relies on several assumptions. The two key challenges when reading Mendelian randomisationterm studies are evaluating the plausibility of the underlying assumptions and interpreting the results. We discuss these challenges below, using terms defined in the glossary ( box 1 ).
Box 1 Glossary of common terms used in Mendelian randomisationterm studies Readers interested in more detailed discussions of these terms are encouraged to read the references. 1 9 12 13 Concepts Instrumental variablesterm·def —variables that are associated with the risk factor of interest, that are not related to confounders, and that affect the outcome only through the risk factor. An instrumental variableterm·def can be any trait (not necessarily a genetic variant) that meets these criteria, but the nature of genetic inheritance means that genetic variants are often plausible instrumental variablesterm·def Mendelian randomisationterm —the use of genetic variants as instrumental variablesterm·def to investigate the effects of modifiable risk factors for disease 6 15 Multiple instruments —the use of more than one genetic variant in a Mendelian randomisationterm analysis Allele score —the number of alleles associated with an increase in the risk factor of interest. These genetic variants are normally identified in large genome-wide association studies. The statistical efficiency (power) of allele scores can be increased by weighting each variant by the size of its association with the risk factor 16 Weak instrument biasterm·def —can occur in Mendelian randomisationterm studies when using one or more genetic variants that only explain a small proportion of the variation in the risk factor, coupled with a small sample size Pleiotropic effects —the effects of a genetic variant on multiple biological pathways. These can either affect the outcome through another trait or pathway to the one under investigation, known as horizontal pleiotropyterm, or affect other traits through the risk factor of interest, known as vertical pleiotropyterm. Horizontal pleiotropyterm is a violation of the instrumental variableterm·def assumptions because the effects of the genetic variant on the outcome are not exclusively through the risk factor; this is problematic for Mendelian randomisationterm studies. Vertical pleiotropyterm is in fact the essence of Mendelian randomisationterm—showing that one factor influences a downstream outcome—and is in general not problematic 1 Statistical methods Single sample Mendelian randomisationterm —using one dataset in the instrumental variableterm·def analysis to yield the causal estimate of the risk factor on the outcome Two sample Mendelian randomisationterm·def —using two different study samples to estimate the instrument-risk factor and instrument-outcome associations to estimate a causal effect of the risk factor on the outcome. This can be useful when the risk factor or outcome, or both, are expensive to measure. 17 18 It also provides an opportunity to substantially increase the statistical power, by incorporating data from multiple sources, including large consortia 17 MR Egger regression —a statistical technique that allows one or more genetic variants to have pleiotropic effects, as long as the size of these pleiotropic effects is independent of the size of the genetic variants’ effects on the risk factor of interest 19 Properties of the genetic instrumentterm·var Testing between observational and Mendelian randomisationterm estimate —tests include the Hausman test (for continuous outcomes in single sample Mendelian randomisationterm) and tests for difference in estimates (for binary outcomes in single sample and two sample Mendelian randomisationterm·def) 20 21 Tests of instrument strength —to evaluate the strength of the association between the instrument (the genetic variant) and the risk factor (for example, the partial F statistic and R 2 ) 22 What assumptions does Mendelian randomisationterm depend on?
Valid instrumental variablesterm are defined by three key assumptionsterm ( table 1 , fig 1 ): that they associate with the risk factor of interest (the relevance assumption); that they share no common cause with the outcome (the independence assumption); and that they do not affect the outcome except through the risk factor (the exclusion restriction assumption). A single genetic variant could plausibly meet these conditions if the biological process linking the variant with the risk factor is well understood. But in many cases Mendelian randomisationterm studies include multiple genetic variants, which can be used in sensitivity analyses to evaluate the underlying assumptions. Generally, the three key assumptionsterm·def must hold for each of the genetic variants. We describe common strategies for assessing the plausibility of these assumptions and give examples, where possible, from published studies.
Table 1 Three key assumptionsterm that must hold for a Mendelian randomisationterm study to be valid Assumption Description Tools to assess plausibility Single sample Two sample Relevance assumption The genetic variants associate with the risk factor of interest The partial F statistic and partial r squared, or risk difference Variants are associated with the risk factor in a large genome-wide study Independence assumption There are no unmeasured confounders of the associations between genetic variants and outcome Covariate balance tests and bias component plots. Adjusting for principal components of population stratification Evidence from large genome-wide association studies on the association of the genetic variants used as instruments with other baseline covariates Exclusion restriction The genetic variants affect the outcome only through their effect on the risk factor of interest Biological knowledge, tests of association of the genetic variants and potential alternative mediating pathways Evidence from large genome-wide association studies that the genetic variants associate with alternative pathways. MR Egger test for pleiotropy, Cook’s distance evaluation of outliers Fig 1 Examples of Mendelian randomisationterm and potential violations of assumptions. (A) A simplified causal diagram depicting confounding of the association of alcohol consumption and blood pressure by existing disease or social deprivation. The instrumental variableterm·def assumptions are that the genetic variants are associated with the risk factor, that theyhave no other influence on the outcome, except through alcohol, and that there are no confounders of the genetic variants-outcome association. (B) Confounding by ancestry could occur if variants associated with alcohol consumption had different frequencies in different ethnic groups in the population sampled and if cultural differences affected blood pressure between ethnic groups. This would violate the second instrumental variableterm·def assumption— the independence assumption. (C) An example of horizontal pleiotropyterm, in which the genetic variants associated with alcohol consumption also affect tobacco consumption (violating the third assumption— the exclusion restriction assumption). (D) An example of vertical pleiotropyterm, in which the effect of ALDH2 on coronary heart disease is mediated by blood pressure. This example does not violate the Mendelian randomisationterm assumptions and does not cause bias.
Sources of data in Mendelian randomisationterm Historically, a typical Mendelian randomisationterm study required measures of genotypes (variants in ALDH2 gene), risk factor (alcohol consumption), and outcome (blood pressure) from the same sample of people. This approach is known as single sample Mendelian randomisationterm. Two sample Mendelian randomisationterm·def involves two separate study populations; for example, data on the ALDH2 genotype and alcohol consumption are measured in one sample, and ALDH2 genotype and blood pressure in the other. 17 23 This design has two advantages. Firstly, neither the risk factor nor the outcome needs to be measured in all studies, which is particularly useful if they are difficult or expensive to measure. Secondly, it allows the summary results from genome-wide association studies to be used, which can be very large (often >50 000) and thus highly precise ( table 2 and table 3 ).
Table 2 Publicly available data sources for two sample Mendelian randomisationterm·def studies Consortium name Description Most recent sample size BCAC 24 Breast cancer 256 123 CARDIoGRAMplusC4D 25 Coronary artery disease and myocardial infarction 184 305 CKDGen 26 Chronic kidney disease 111 666 DIAGRAM 27 Diabetes 159 208 EAGLE 28 Antenatal and early life and childhood phenotypes 47 541 EGG 29 Early growth 153 781 GIANT 30 Height, BMI, and other adiposity traits 693 529 GLGC 31 Global lipids genetics consortium 331 368 ISGC 32 Stroke 84 961 MAGIC 33 Glucose and insulin related traits 224 459 PGC 34 35 Psychiatric genetics, alcohol and tobacco, and other related traits >500 000 SSGAC 36 Educational attainment and wellbeing 293 723 Table 3 Databases of genome-wide association study results Data source Description Number of traits Integrated with statistics package?
MR-Base A curated database of genome-wide association study results with integrated R package for MR 23 Over 1000 Yes PhenoScanner A curated database of genome-wide association study results with integrated R package for MR 37 Over 500 Yes GWAS catalog Searchable database of genome-wide association study results 38 Over 24 000 No Statistical power is usually much higher in two sample studies. These advantages come with two additional assumptions: the two samples are assumed to represent the same underlying population, and overlap in participantsterm·var between the two samples can cause bias towards the risk factor-outcome association. 17 39 Single or multiple genetic variants The simplest applications of Mendelian randomisationterm use a single genetic variant as an instrument for the risk factor. Some of the most persuasive published studies use a single genetic variant with a relatively well understood function, so the core assumptions can be supported by biological knowledge. One such example is the use of a missense variant in the PCSK9 gene, which modifies the function of proprotein convertase subtilisin/kexin type 9, an enzyme that degrades the low density lipoprotein (LDL) receptor on hepatocytes. PCSK9 variants are associated with altered blood concentrations of LDL cholesterol and risk of heart disease. 40 41 This provides further evidence that LDL cholesterol causes heart disease, and also indicates that drugs that inhibit PCSK9 may have cardiovascular benefits. This was confirmed in phase III clinical trials. 42 43 In most circumstances, however, single genetic variants individually typically explain only a very small proportion of the variation in a phenotype; investigators may refer to these as “weak instruments,” particularly in modest sample sizes. Studies using variants with modest effects in small samples are likely to have very low statistical power and can be biased. To overcome this, investigators have developed methods that use multiple genetic variants that collectively explain more of the variation in a risk factor than a single variant and thus have more statistical power. One way of using multiple genetic variants is to use single nucleotide polymorphisms (SNPs, the most common form of DNA variation among people) as individual instruments in a statistical regression model in the setting of a single or two sample dataset (multiple instruments approach). The second way is to aggregate the variants into an allele score (also called a genetic risk score, gene score, or a multilocus allele score). The allele score is effectively a single instrumental variableterm·def that can be used to predict the risk factor in a Mendelian randomisationterm analysis.
Genetic variants in these scores are often weighted by their associations with the risk factor to maximise statistical power. Many recent Mendelian randomisationterm investigations of complex traits (such as blood pressure, BMI, or blood lipids) 44 have used multiple variants because they can help identify pleiotropy or other violations of the underlying assumptions.
Genetic pleiotropy Genetic variants may affect the outcome through pathways other than through the risk factor of interest (so called horizontal pleiotropic effects, fig 1C ). Genetic variants associated with alcohol consumption, for example, may affect other behaviours such as smoking, which would invalidate the exclusion restriction. When using single or multiple genetic variants, Mendelian randomisationterm estimates require that genetic variants do not have such horizontally pleiotropic effects. This means the results could be biased if a genetic variant or an allele score has pleiotropic effects on the outcome that are not mediated through the risk factor of interest ( fig 2 ). Genetic variants can also affect the outcome through a pathway affected by the risk factor of interest (vertical pleiotropic effects, fig 1D ). This does not invalidate the instrumental variableterm·def assumptions and does not result in bias.
Fig 2 Example of genetic pleiotropy in Mendelian randomisationterm: HDL cholesterol and risk of heart disease. Variants associated with HDL cholesterol are likely to have pleiotropic effects on risk of heart disease because they also associate with LDL cholesterol and triglycerides. 2 45 Thus the inverse variance weighted Mendelian randomisationterm estimate, which assumes no pleiotropy, provides (biased) evidence of a protective role for HDL cholesterol in coronary heart disease. But the estimates using MR Egger, weighted median, and weighted mode, which allow for genetic pleiotropy, are attenuated towards the null. The MR Egger estimator assumes that for the variants with pleiotropic effects on coronary heart disease the magnitude of these effects do not correlate with the magnitude of the variants’ effects on HDL cholesterol. These results suggest that the inverse variance weighted estimate is driven by genetic pleiotropy 19 46 47 48 and that HDL cholesterol is unlikely to have a major causal role in the development of coronary heart disease. CHD=coronary heart disease; ERFC=Emerging Risk Factors Collaboration; HDL-C=high density lipoprotein cholesterol; SD=standard deviation Various methods have been developed that allow for genetic pleiotropy. 19 46 47 These provide useful sensitivity analyses to explore whether a finding depends on the assumption that all the variants have no pleiotropic effects. One such approach, known as the median estimator, can provide reliable evidence as long as at least half the genetic variants have no pleiotropic effects. 47 49 A second method, known as MR Egger regression, allows all variants to have pleiotropic effects, provided they are not proportional to the variants’ effects on the risk factor of interest. 19 A comprehensive analysis of cross trait genetic effects indicates that the magnitude of genome-wide associations are generally independent across traits that are not causally related, so the assumption may be plausible for many potential relationships. 50 MR Egger regression yields less precise estimates than other methods, owing to a power penalty. Most of these methods assume that the risk factor has the same effect on everyone.
How to assess the key assumptions Tests of instrument strength—the relevance assumption The power of a Mendelian randomisationterm study is determined by sample size and strength of the association between the proposed instrument and risk factor. Weak instruments that poorly predict the risk factor cause three problems. Firstly, they provide very little statistical power to test hypotheses. Secondly, bias due to violations of the core instrumental variableterm·def assumptions, such as horizontally pleiotropic effects of variants, will be amplified. Thirdly, even when using very large samples, results using weak instruments are biased towards the outcome-risk factor association in the single sample setting and towards the null in the two sample setting. Precision (assessed with confidence intervals) is underestimated. Weak instruments can be detected using the F statistic for single sample settings. 51 A rule of thumb is that the F statistic should be greater than 10. Exceeding this threshold indicates that a result based on a valid instrumental variableterm ought not to suffer substantially from weak instrument biasterm·def but does not guarantee sufficient statistical power to test a specific hypothesis.
Independence and exclusion restriction assumptions Confounding by violations of the independence assumption (no confounders) and exclusion restriction (genetic instruments work only through the risk factor) can be investigated by estimating the relation between the genetic instrumentterm·var and a wide range of characteristics, analogous to the balance of baseline characteristics between treatment arms in a randomised trial. 52 53 54 Although these assessments cannot prove that the independence and exclusion restriction assumptions hold, they can provide evidence regarding their (lack of) plausibility.
One promising way to evaluate the plausibility of the assumptions is using “negative control” populations. Genetic variants that are known to affect alcohol consumption, for example, should not be associated with outcomes in populations that rarely or never drink, such as children or women in some societies. This tests the exclusion restriction and independence assumptions. If investigators found an association between ALDH2 variants and blood pressure in a non-drinking population, the instrumental variableterm·def assumptions may not hold in the original population either. The usefulness of negative controls depends on them being well reasoned and, of course, available.
How are Mendelian randomisationterm studies analysed?
The simplest approach is to report the association of the genetic variants and the outcome. This test is relatively robust and should be reported in all studies. It provides evidence as to whether the risk factor causes the outcome, but it is not informative about the size of the effect. The magnitude of the causal effect can be estimated by dividing the genetic variant-outcome association by the genetic variant-risk factor association. This ratio is known as the instrumental variableterm·def or Wald estimate. 55 Inclusion of, or stratifying by, the risk factor of interest To assess whether a genetic variant (or several in combination) mediates its effect on an outcome through the primary risk factor (and thus to test for exclusion restriction), some studies adjust or stratify for the primary risk factor of interest. But this is problematic for several reasons. Firstly, a residual association of the genetic instrumentterm·var with the disease after adjustment for the risk factor does not imply that the variant affects the disease through other pathways. The measured risk factor almost certainly does not completely account for the lifetime variation in the trait instrumented by the gene. A residual effect of the gene (on the outcome after adjusting for the measured risk factor) might therefore not indicate a violation of the exclusion restriction assumption. Secondly, adjusting or stratifying for the risk factor can lead to collider bias (see supplementary figure 1), in the same way that conducting a subgroup analysis by “on-treatment” response breaks the randomisation in a trial. 57 To see why, consider a study investigating the effect of statins on coronary heart disease. 58 The on-treatment (or “achieved”) LDL cholesterol concentration will break the randomisation and is likely to be biased by other characteristics that influence lipid levels (such as age, sex, diet, physical activity, and so on). Although this is a seemingly common retrospective approach to try and elucidate dose-response relatonships in a randomised trial, such analyses can be as vulnerable to confounding as observational analyses. Similarly, adjusting (or stratifying) a Mendelian randomisationterm analysis for the risk factor can induce spurious correlations between the genetic variants and the outcome.
Why do some Mendelian randomisationterm studies adjust for other traits?
If a genetic variant is a valid instrument, inclusion of other covariates is not necessary, but they can increase statistical efficiency. Some Mendelian randomisationterm studies adjust for other characteristics, including directly measured traits (such as age, sex, or smoking) or associations between the genetic instrumentterm·var and traits other than the risk factor of interest. The genetic variant-risk factor and genetic variant-outcome associations should be adjusted for the same covariates. If not, the instrumental variableterm·def estimate may be biased. For example, if a two sample Mendelian randomisationterm·def study of the effect of BMI on blood pressure used a blood pressure genome-wide association study that adjusted for BMI, then the estimate would be unreliable and the direction of effect could even be reversed. These problems are an area of active methodological research.
In some samples, the association between a genetic variant and outcome may be confounded by hidden population structure. This can be tackled by adjusting for genetic ancestry or restricting to ethnically homogenous samples. Suppose a study sampled data from an ethnically mixed east Asian and European population. The minor allele of the alcohol variant rs671 is extremely rare in European populations, so an association between this variant and an outcome in this sample could be due to differences in ethnicity ( fig 1B ). This problem could be mitigated by stratifying by ethnicity or adjusting for genetic principal components (which are calculated in datasets with genome-wide arrays and provide proxy measures of genetic ancestry). The UK Biobank, for example, provides up to 40 principal components for adjusting genetic analyses for differences in ethnicity. This improves reliability, as has been done in a recent Mendelian randomisationterm study of adiposity and risk of heart disease and diabetes. If a study depends on these adjustments—that is, there are major differences between the unadjusted and adjusted findings—the results should be treated with caution. At the very least the source of potential bias must be investigated and justified, just as for residual confoundingterm in a conventional epidemiology study.
Why do some Mendelian randomisationterm studies remove variants from the genetic instrumentterm·var?
Some studies manually remove (or “prune”) genetic variants thought to be pleiotropic from a genetic instrument to provide a more reliable estimate of the association. Voight and colleagues used genetic variants that were exclusively associated with HDL cholesterol (not LDL cholesterol or triglycerides) to investigate the relationship between HDL cholesterol and heart disease. 60 Including or excluding a genetic variant in this way can be arbitrary and limited by available data; such pruning should be done with caution (if at all) and for well justified and transparent reasons. Furthermore, as the sample size increases, genetic variants will be identified as showing associations with multiple traits owing to vertical or horizontal pleiotropyterm. Excluding variants on the basis of vertical pleiotropic effects will bias findings, since these effects would be ones seen when manipulating the risk factor of interest. One or more variants might be excluded from sensitivity analyses if they are obvious outliers based on visual assessment of the data or using more formal approaches such as Cook’s distance. 63 How are Mendelian randomisationterm studies reported?
Having checked whether the assumptions underpinning a Mendelian randomisationterm analysis are valid ( table 1 ), readers should be familiar with how such studies are reported ( box 2 ). They may report the association of one or more genetic variants with the outcome or provide an estimate of the causal (instrumental variableterm·def) effects of the risk factor on the outcome, or both. Ference and colleagues report the association between SNPs that alter LDL cholesterol and risk of heart disease, showing compelling genetic evidence of a dose-response association across multiple independent genetic loci. These SNPs can also be used to provide an “overall” causal effect of LDL cholesterol on the risk of coronary heart disease (odds ratio 0.46, 95% confidence ratio 0.41 to 0.51 per 1mmol/L lower LDL cholesterol). These associations leverage genetic differences that occur at conception and can detect whether differences in the risk factor at any point over the life course affect the outcome. As a guide to future research and drug development, it is important to remember that findings of Mendelian randomisationterm potentially reflect lifetime differences in risk factors. 65 66 Box 2 Critical appraisal checklist for evaluating Mendelian randomisationterm studies Some key questions readers can ask below.
Core Mendelian randomisationterm assumptions Is there sufficient evidence that the genetic variants are robustly associated with the risk factor of interest?
Are the genetic variants associated with potential confounders? Do the authors present this relationship?
Is there any way for the genetic variants to affect the outcome through alternative pathways (horizontal pleiotropyterm)? Do the authors present alternative Mendelian randomisationterm approaches (such as MR Egger, median, and mode estimators, or use of “negative control” populations) to investigate this more fully?
Methods reporting All studies Are the effect and other alleles coded in the same direction for the exposure and outcome?
Two sample studies Were the two samples drawn from the same population?
Were the two samples independent?
Was the analysis restricted to independent variants (that is, pruned of SNPs in linkage disequilibrium) or did the analysis allow for the correlation between variants?
Data presentation Do the authors present the results as a genetic association, an instrumental variableterm·def estimate, or both?
If they provide an instrumental variableterm·def estimate, do they compare it with the conventional observational estimate?
Do the authors provide sensitivity analyses such as MR Egger, weighted median, and mode Mendelian randomisationterm, or use negative control populations?
Do the authors manually pick and choose which SNPs go into the instrument to tackle pleiotropy? If so, is the approach and justification clear?
Do the authors provide the data that they used (especially for Mendelian randomisationterm analyses conducted at the summary level) in a supplement to allow researchers to reproduce their findings?
Interpretation If the Mendelian randomisation estimateterm is similar to the observational estimate and provides evidence in support of a causal effect, could it be due to weak instrument biasterm·def in a single study or confounding through, for example, horizontal pleiotropyterm?
If the Mendelian randomisation estimateterm differs from the observational estimate and provides little evidence of a causal effect, could this be due to weak instrument biasterm·def when using two different samples or negative confounding due to pleiotropy?
Mendelian randomisationterm provides estimates of the effects of the risk factor over a lifetime, and the numerical effect estimates may not be clinically meaningful. Will interventions at a specific age have the same sized effects?
Are the 95% confidence intervals of the Mendelian randomisation estimateterm sufficiently precise to identify the observational estimate and a clinically meaningful difference?
Clinical implications Do the results triangulate with other forms of evidence? Could a clinical trial be conducted to provide definitive evidence, as in the case of PCSK9 inhibitors?
If a randomised clinical trial is not feasible (such as in the case of alcohol consumption and risk of heart disease) or unlikely to be conducted in the short term (such as the case of lifestyle interventions to lower BMI and risk of heart disease), and there is existing evidence from multiple Mendelian randomisationterm studies and other robust study designs that converge on a similar result and show consistency of association, this information can be used to guide patient care; for example, advising weight loss to prevent heart disease or advising against moderate alcohol consumption to prevent cardiovascular disease Special cases of reporting findings Testing for differences between Mendelian randomisationterm estimates from different SNPs If multiple genetic variants and biological pathways influence the risk factor, we can test whether the effect of the risk factor on the outcome is similar when using different variants using Hansen tests or Cochrane’s Q test for individual level data and summary data, respectively. 67 68 If these test statistics are large (yielding correspondingly small P values), the estimated causal effects of the risk factor may vary across the population or between variants. This might be due to multiple pathways causing the outcome. If these tests cannot be rejected, sample size and statistical power may be insufficient to detect differences that do exist and, therefore, may be falsely reassuring.
Comparing findings from observational and Mendelian randomisationterm analyses in the same dataset For continuous outcomes in single sample analyses, the Hausman test can be used to assess whether the Mendelian randomisationterm and linear regression resultsterm (obtained from the same dataset) are systematically different. Differences could occur because the linear regression resultsterm are biased by residual confoundingterm, the independence or exclusion restriction assumptions underlying the instrumental variableterm·def regression approach are invalid, or both. Finally, the effects being estimated by the two methods may not be the same—the Mendelian randomisation estimateterm reflects the effects of lifelong perturbations in the risk factor, whereas linear regression resultsterm may reflect more acute effects. Mendelian randomisationterm estimates are almost always less precise and have wider confidence intervals than linear regression ( fig 3 ), so tests for difference often have low statistical power. Fig 3 Example associations between risk factors and outcomes from traditional observational epidemiology and Mendelian randomisationterm instrumental variableterm·def estimates. 60 69 70 71 72 73 For some associations—such as vitamin D and mortality—the Mendelian randomisationterm results potentially confirm some causal relation. For other associations—CRP and heart disease—the Mendelian randomisationterm results are consistent with there being no causal effect. BMI=body mass index; CRP=C reactive protein; HDL-C=high density lipoprotein cholesterol; LDL-C=low density lipoprotein cholesterol How should findings from Mendelian randomisationterm be interpreted?
The findings from Mendelian randomisationterm studies, which are less susceptible to confounding and reverse causality bias, sit at the interface between traditional observational epidemiology and interventional trials ( fig 4 ). A well conducted Mendelian randomisationterm study that reasonably satisfies the above assumptions often provides more reliable evidence than a conventional observational study. But the findings should be interpreted in the context of existing evidence from other sources, using different study designs, 75 and clinical guidelines should not be rewritten solely on the basis of Mendelian randomisationterm results.
Fig 4 A hierarchy of observational and experimental data. Mendelian randomisationterm studies sit at the interface of experimental and observational studies. Their findings can be used to provide more reliable evidence to guide interventional research and provide information about potential public health interventions when a randomised controlled trial may not be feasible. Although we adapt the conventional pyramid of evidence for presentation purposes, we consider that triangulation of findings from different study designs should be used. Triangulating the evidence Some studies report the Mendelian randomisation estimateterm in the context of traditional observational epidemiology ( fig 3 ; supplementary table 1). When there is clear evidence that the associations differ, the Mendelian randomisation estimateterm can provide strong evidence that the traditional observational estimate arises from confounding and/or reverse causality. In the case of C-reactive protein and coronary heart disease, the Mendelian randomisation estimatesterm suggest that targeting CRP is unlikely to be a viable therapeutic target for the prevention of coronary heart disease. 72 In contrast, when the Mendelian randomisation estimateterm has very wide confidence intervals that overlap the observational estimate and include the null (as in the case of dairy consumption and blood pressure), 69 very little can be inferred from the Mendelian randomisationterm results.
Summary estimates from observational and Mendelian randomisationterm studies can be compared formally using tests for heterogeneity, such as Cochrane’s Q test and tests for differences in estimates. This can provide statistical evidence of a difference, but whether this should be interpreted as clinically meaningful is further outlined below.
Clinical and public health implications Clinical and public health decisions about potential interventions ideally require evidence about the size and direction of an effect for a specific population. We might want to know, for example, how reducing alcohol consumption by an average of one serving a day beginning at age 45 affects blood pressure five years later. But this must be approached with care because genetic variants generally relate to lifelong differences in a risk factor, not the effects at a specific time. ALDH2 is associated with alcohol consumption from adolescence through middle age, 76 so the Mendelian randomisationterm results (coupled with biological knowledge) indicate that consuming less alcohol would lead to lower blood pressure on average, but do not necessarily inform the effect size of an intervention at a specific time in life for a specific duration. Another example is the relation between vitamin D and multiple sclerosis, where data from Mendelian randomisationterm studies indicate a protective effect, but such protection may be limited to childhood and adolescence and not later adult life. 66 77 Biological and methodological knowledge about the relations between risk factors and outcomes is critical to interpreting Mendelian randomisationterm studies. Does risk accumulate over time? Does the risk factor have an acute effect? Is there a particular time in life where this risk factor has an effect on this disease? 6 66 78 79 This is why some studies focus on the association of the genetic variants and outcome, 56 rather than providing a direct quantification of the causal effect scaled to a difference in the exposure of interest.
Consider another example, where the biology and context are important for the appraisal of Mendelian randomisationterm findings. Variants in the CHRNA5 geneterm that associate with increased heaviness of smoking are (through smoking behaviour) associated with increased risk of lung cancerterm. People with these variants are also likely to find it more difficult to reduce smoking after diagnosis of lung cancer. But variants in this gene cannot be used to assess the effect of quitting or reducing smoking on mortality after diagnosis with lung cancer, because people with the risk variants will have smoked more heavily before diagnosis. The association between CHRNA5 risk variants and mortality in people with lung cancer reflects an individual’s lifetime exposure to tobacco smoke, and we would not expect an instrumental variableterm·def estimate using CHRNA5 to reflect the effect of current smoking on prognosis. Most importantly, the triggers of disease onsetterm may be entirely different to the factors leading to progressionterm. This means that Mendelian randomisationterm studies need to be performed using data from studies of disease progression and prognosis, rather than disease occurrence, to generate findings of direct relevance to treating disease. 81 Finally, when an exposure is seemingly non-modifiable (such as height 82 83 ), a causal relation can point towards potential mechanisms or pathways that lead to disease (even if the actual risk factor under investigation is non-modifiable); such pathways are likely to be modifiable, providing new insights in disease aetiology.
Summary Mendelian randomisationterm studies can provide reliable evidence on the effect of modifiable risk factors for disease or ill health and can overcome some limitations of traditional observational epidemiology. In settings for which the instrumental variableterm·def assumptions are well justified (assessed as described above and using biological knowledge), the findings could help prioritise clinical trials or drug development and inform clinical or public health decision making. 84 85 86 The conversation on how to conduct, report, and interpret Mendelian randomisationterm studies is still ongoing. Other areas of medical research, such as randomised controlled trials, have been greatly strengthened by close collaborations between methodologists, empirical researchers, and clinicians. Similar collaborations are needed to ensure the strengths and limitations of Mendelian randomisationterm are fully appreciated and realised.
We thank Sonja Swanson (Erasmus University Rotterdam) for contributing to earlier versions of the manuscript. We are additionally grateful to Gene Feder and David Kessler (University of Bristol), Iain Turnbull (University of Oxford), Tony Ryan (GP, Portishead) and Philip Pucher (surgeon, Imperial College London) for valuable feedback on the manuscript.
Web extra Extra material supplied by authors Supplementary table: Strength of genetic instruments Supplementary figure: An illustration of collider bias Contributors and sources: All authors wrote and edit the manuscript, and approved the final draft. All authors confirm that they have contributed to this article and met the following three requirements: (a) they made significant contributions to the conception and design; (b) they drafted or revised the article for intellectual content; and (c) gave final approval of the published article. NMD is the guarantor.
Funding: The Medical Research Council (MRC) and the University of Bristol fund the MRC Integrative Epidemiology Unit (MC_UU_12013/1, MC_UU_12013/9). NMD is supported by the Economics and Social Research Council via a Future Research Leaders Fellowship (ES/N000757/1). MVH works in a unit that receives funding from the MRC and is supported by a British Heart Foundation Intermediate Clinical Research Fellowship (FS/18/23/33512) and the National Institute for Health Research Oxford Biomedical Research Centre. No funding body has influenced data collection, analysis, or its interpretations.
Competing interests: We have read and understood BMJ policy on declaration of interests and declare that we have no competing interests. Davey SmithpersName G Hemani G Mendelian randomization: genetic anchors for causal inference in epidemiological studies . Hum Mol Genet 2014 ; 23 ( R1 ): R89 - 98 . 10.1093/hmg/ddu328 25064373 PMC4170722 2 Burgess S Small DS Thompson SG A reviewterm·mth of instrumental variableterm·def estimators for Mendelian randomization . Stat Methods Med Res Published Online First 2015 ; 17 . 10.1177/0962280215597579 . 26282889 PMC5642006 3 Haycock PC Burgess S Wade KH Bowden J Relton C Davey SmithpersName G Best (but oft-forgotten) practices: the design, analysis, and interpretation of Mendelian randomization studies . Am J Clin Nutr 2016 ; 103 : 965 - 78 . 10.3945/ajcn.115.118216 26961927 PMC4807699 4 DaviespersName NM Davey SmithpersName G Windmeijer F Martin RM Issues in the reporting and conduct of instrumental variableterm·def studies: a systematic reviewterm·mth . Epidemiology 2013 ; 24 : 363 - 9 . 10.1097/EDE.0b013e31828abafb 23532055 5 Swanson SA Hernán MA Commentary: how to report instrumental variableterm·def analyses (suggestions welcome) . Epidemiology 2013 ; 24 : 370 - 4 . 10.1097/EDE.0b013e31828d0590 23549180 6 Davey SmithpersName G Ebrahim S ‘Mendelian randomization’: can genetic epidemiology contribute to understanding environmental determinants of disease? Int J Epidemiol 2003 ; 32 : 1 - 22 . 10.1093/ije/dyg070 12689998 7 Davey SmithpersName G Ebrahim S What can mendelian randomisationterm tell us about modifiable behavioural and environmental exposures? BMJ 2005 ; 330 : 1076 - 9 . 10.1136/bmj.330.7499.1076 15879400 PMC557238 8 Evans DM Davey SmithpersName G Mendelian randomization: new applications in the coming age of hypothesis-free causality . Annu Rev Genomics Hum Genet 2015 ; 16 : 327 - 50 . 10.1146/annurev-genom-090314-050016 25939054 9 Burgess S Butterworth A Malarstig A Thompson SG Use of Mendelian randomisationterm to assess potential benefit of clinical intervention . BMJ 2012 ; 345 : e7325 - 7325 .. 10.1136/bmj.e7325 23131671 10 Takagi S Baba S Iwai N The aldehyde dehydrogenase 2 gene is a risk factor for hypertension in Japanese but does not alter the sensitivity to pressor effects of alcohol: the Suita study . Hypertens Res 2001 ; 24 : 365 - 70 . 10.1291/hypres.24.365 11510748 11 Chen L Davey SmithpersName G Harbord RM Lewis SJ Alcohol intake and blood pressure: a systematic reviewterm·mth implementing a Mendelian randomization approach . PLoS Med 2008 ; 5 : e52 . 10.1371/journal.pmed.0050052 18318597 PMC2265305 12 Hernán M Robins J Causal Inference. Chapman & Hall/CRC , 2018 , ( forthcoming ). Burgess S, Thompson SG. Mendelian randomization: methods for using genetic variants in causal estimation . Boca Raton: CRC Press, Taylor & Francis Group 2015. Imbens GW Angrist JD Identification and estimation of local average treatment effects . Econometrica 1994 ; 62 : 467 10.2307/2951620 . Lawlor DA Harbord RM Sterne JAC Timpson N Davey SmithpersName G Mendelian randomization: using genes as instruments for making causal inferences in epidemiology . Stat Med 2008 ; 27 : 1133 - 63 . 10.1002/sim.3034 17886233 16 Burgess S Thompson SG Use of allele scores as instrumental variablesterm·def for Mendelian randomization . Int J Epidemiol 2013 ; 42 : 1134 - 44 . 10.1093/ije/dyt093 24062299 PMC3780999 17 Pierce BL Burgess S Efficient design for Mendelian randomization studies: subsample and 2-sample instrumental variableterm·def estimators . Am J Epidemiol 2013 ; 178 : 1177 - 84 . 10.1093/aje/kwt084 23863760 PMC3783091 18 Inoue A Solon G Two-sample instrumental variablesterm·def estimators . Rev Econ Stat 2010 ; 92 : 557 - 61 . 10.1162/REST_a_00011 . Bowden J Davey SmithpersName G Burgess S Mendelian randomization with invalid instruments: effect estimation and bias detection through Egger regression . Int J Epidemiol 2015 ; 44 : 512 - 25 . 10.1093/ije/dyv080 26050253 PMC4469799 20 Hausman JA Specification tests in econometrics . Econometrica 1978 ; 46 : 1251 - 71 . 10.2307/1913827 . Altman DG Bland JM Interaction revisited: the difference between two estimates . BMJ 2003 ; 326 : 219 . 10.1136/bmj.326.7382.219 12543843 PMC1125071 22 Angrist JD Pischke JS Mostly harmless econometrics: an empiricist’s companion. Princeton Univ Pr , 2009 . Hemani G Zheng J Wade KH The MR-base enables systematic causal inference across the phenome . eLife [forthcoming] . 10.7554/eLife.34408 PMC5976434 29846171 24 Michailidou K Lindström S Dennis J NBCS Collaborators ABCTB Investigators ConFab/AOCS Investigators Association analysis identifies 65 new breast cancer risk loci . Nature 2017 ; 551 : 92 - 4 . 10.1038/nature24284 29059683 PMC5798588 25 Nikpay M Goel A Won HH A comprehensive 1,000 Genomes-based genome-wide association meta-analysis of coronary artery disease . Nat Genet 2015 ; 47 : 1121 - 30 . 10.1038/ng.3396 26343387 PMC4589895 26 Li M Li Y Weeks O CHARGE Glycemic-T2D Working Group CHARGE Blood Pressure Working Group SOS2 and acp1 loci identified through large-scale exome chip analysis regulate kidney development and function . J Am Soc Nephrol 2017 ; 28 : 981 - 94 . 10.1681/ASN.2016020131 27920155 PMC5328154 27 Scott RA Scott LJ Mägi R ; DIAbetes Genetics Replication And Meta-analysis (DIAGRAM) Consortium. an expanded genome-wide association study of type 2 diabetes in Europeans . Diabetes 2017 ; 66 : 2888 - 902 . 10.2337/db16-1253 28566273 PMC5652602 28 Felix JF Bradfield JP Monnereau C Bone Mineral Density in Childhood Study (BMDCS) Early Genetics and Lifecourse Epidemiology (EAGLE) consortium Early Growth Genetics (EGG) Consortium Bone Mineral Density in Childhood Study BMDCS Genome-wide association analysis identifies three new susceptibility loci for childhood body mass index . Hum Mol Genet 2016 ; 25 : 389 - 403 . 10.1093/hmg/ddv472 26604143 PMC4854022 29 Horikoshi M Beaumont RN Day FR CHARGE Consortium Hematology Working Group Genome-wide associations for birth weight and correlations with adult disease . Nature 2016 ; 538 : 248 - 52 . 10.1038/nature19806 27680694 PMC5164934 30 Yengo L Sidorenko J Kemper KE Meta-analysis of genome-wide association studies for height and body mass index in ~700,000 individuals of European ancestry . bioRxiv 2018 , 10.1101/274654 . PMC6488973 30124842 31 Liu DJ Peloso GM Yu H Charge Diabetes Working Group EPIC-InterAct Consortium EPIC-CVD Consortium GOLD Consortium VA Million Veteran Program Exome-wide association study of plasma lipids in >300,000 individuals . Nat Genet 2017 ; 49 : 1758 - 66 . 10.1038/ng.3977 29083408 PMC5709146 32 Neurology Working Group of the Cohorts for Heart and Aging Research in Genomic Epidemiology (CHARGE) Consortium, the Stroke Genetics Network (SiGN), and the International Stroke Genetics Consortium (ISGC) Identification of additional risk loci for stroke and small vessel disease: a meta-analysis of genome-wide association studies . Lancet Neurol 2016 ; 15 : 695 - 707 . 10.1016/S1474-4422(16)00102-2 27068588 PMC4943223 33 Shungin D Winkler TW Croteau-Chonka DC ADIPOGen Consortium CARDIOGRAMplusC4D Consortium CKDGen Consortium GEFOS Consortium GENIE Consortium GLGC ICBP International Endogene Consortium LifeLines Cohort Study MAGIC Investigators MuTHER Consortium PAGE Consortium ReproGen Consortium New genetic loci link adipose and insulin biology to body fat distribution . Nature 2015 ; 518 : 187 - 96 . 10.1038/nature14132 25673412 PMC4338562 34 Wray NR Ripke S Mattheisen M Genome-wide association analyses identify 44 risk variants and refine the genetic architecture of major depression . Nat Genet 2018 ; 50 : 668 - 81 . 29700475 10.1038/s41588-018-0090-3 PMC5934326 35 Sullivan PF Agrawal A Bulik CM Psychiatric Genomics Consortium Psychiatric genomics: an update and an agenda . Am J Psychiatry 2018 ; 175 : 15 - 27 . 10.1176/appi.ajp.2017.17030283 28969442 PMC5756100 36 Okbay A Beauchamp JP Fontana MA LifeLines Cohort Study Genome-wide association study identifies 74 loci associated with educational attainment . Nature 2016 ; 533 : 539 - 42 . 10.1038/nature17671 27225129 PMC4883595 37 Staley JR Blackshaw J Kamat MA PhenoScanner: a database of human genotype-phenotype associations . Bioinformatics 2016 ; 32 : 3207 - 9 .. 10.1093/bioinformatics/btw373 27318201 PMC5048068 38 MacArthur J Bowler E Cerezo M The new NHGRI-EBI Catalog of published genome-wide association studies (GWAS Catalog) . Nucleic Acids Res 2017 ; 45 ( D1 ): D896 - 901 . 10.1093/nar/gkw1133 27899670 PMC5210590 39 Burgess S DaviespersName NM Thompson SG Bias due to participant overlap in two-sample Mendelian randomization . Genet Epidemiol 2016 ; 40 : 597 - 608 . 10.1002/gepi.21998 27625185 PMC5082560 40 Cohen JC Boerwinkle E Mosley TH Jr Hobbs HH Sequence variations in PCSK9, low LDL, and protection against coronary heart disease . N Engl J Med 2006 ; 354 : 1264 - 72 . 10.1056/NEJMoa054013 16554528 41 Abifadel M Varret M Rabès J-P Mutations in PCSK9 cause autosomal dominant hypercholesterolemia . Nat Genet 2003 ; 34 : 154 - 6 . 10.1038/ng1161 12730697 42 Sabatine MS Giugliano RP Keech AC FOURIER Steering Committee and Investigators Evolocumab and clinical outcomes in patients with cardiovascular disease . N Engl J Med 2017 ; 376 : 1713 - 22 . 10.1056/NEJMoa1615664 28304224 43 American College of Cardiology. ODYSSEY outcomes: results suggest use of PCSK9 inhibitor reduces CV events, LDL-C in ACS patients. Mar 2018. http://www.acc.org/latest-in-cardiology/articles/2018/03/05/15/53/sat-9am-odyssey-outcomes-cv-outcomes-with-alirocumab-after-acs-acc-2018 44 Benn M Nordestgaard BG From genome-wide association studies to Mendelian randomization: Novel opportunities for understanding cardiovascular disease causality, pathogenesis, prevention, and treatment . Cardiovasc Res 2018 . 10.1093/cvr/cvy045 29471399 45 HolmespersName MV Asselbergs FW Palmer TM UCLEB consortium Mendelian randomization of blood lipids for coronary heart disease . Eur Heart J 2015 ; 36 : 539 - 50 . 10.1093/eurheartj/eht571 24474739 PMC4344957 46 Hartwig FP Davey SmithpersName G Bowden J Robust inference in summary data Mendelian randomization via the zero modal pleiotropy assumption . Int J Epidemiol 2017 ; 46 : 1985 - 98 . 10.1093/ije/dyx102 29040600 PMC5837715 47 Bowden J Davey SmithpersName G Haycock PC Burgess S consistent estimation in mendelian randomization with some invalid instruments using a weighted median estimator . Genet Epidemiol 2016 ; 40 : 304 - 14 . 10.1002/gepi.21965 27061298 PMC4849733 48 Di Angelantonio E Sarwar N Perry P Emerging Risk Factors Collaboration Major lipids, apolipoproteins, and risk of vascular disease . JAMA 2009 ; 302 : 1993 - 2000 . 10.1001/jama.2009.1619 19903920 PMC3284229 49 Kang H Zhang A Cai TT Instrumental variablesterm·def estimation with some invalid instruments and its application to mendelian randomization . J Am Stat Assoc 2016 ; 111 : 132 - 44 . 10.1080/01621459.2014.994705 . Pickrell JK Berisa T Liu JZ Ségurel L Tung JY Hinds DA Detection and interpretation of shared genetic influences on 42 human traits . Nat Genet 2016 ; 48 : 709 - 17 . 10.1038/ng.3570 27182965 PMC5207801 51 Staiger D Stock J Instrumental variablesterm·def regression with weak instruments . Econometrica 1997 ; 65 : 557 - 86 . 10.2307/2171753 . Davey SmithpersName G Lawlor DA Harbord R Timpson N Day I Ebrahim S Clustered environments and randomized genes: a fundamental distinction between conventional and genetic epidemiology . PLoS Med 2007 ; 4 : e352 . 10.1371/journal.pmed.0040352 18076282 PMC2121108 53 Tchetgen Tchetgen EJ Walter S Glymour MM Commentary: building an evidence base for mendelian randomization studies: assessing the validity and strength of proposed genetic instrumental variablesterm·def . Int J Epidemiol 2013 ; 42 : 328 - 31 . 10.1093/ije/dyt023 23508417 PMC3600630 54 Glymour MM Tchetgen Tchetgen EJ Robins JM Credible Mendelian randomization studies: approaches for evaluating the instrumental variableterm·def assumptions . Am J Epidemiol 2012 ; 175 : 332 - 9 . 10.1093/aje/kwr323 22247045 PMC3366596 55 Boef AGC Dekkers OM le Cessie S Mendelian randomization studies: a reviewterm·mth of the approaches used and the quality of reporting . Int J Epidemiol 2015 ; 44 : 496 - 511 . 10.1093/ije/dyv071 25953784 56 HolmespersName MV Dale CE Zuccolo L InterAct Consortium Association between alcohol and cardiovascular disease: Mendelian randomisationterm analysis based on individual participant data . BMJ 2014 ; 349 : g4164 - 4164 . 10.1136/bmj.g4164 25011450 PMC4091648 57 Rosenbaum PR The consquences of adjustment for a concomitant variable that has been affected by the treatment . J R Stat Soc Ser Gen 1984 ; 147 : 656 10.2307/2981697 . Giugliano RP Pedersen TR Park J-G FOURIER Investigators Clinical efficacy and safety of achieving very low LDL-cholesterol concentrations with the PCSK9 inhibitor evolocumab: a prespecified secondary analysis of the FOURIER trial . Lancet 2017 ; 390 : 1962 - 71 .. 10.1016/S0140-6736(17)32290-0 28859947 59 Lyall DM Celis-Morales C Ward J Association of body mass index with cardiometabolic disease in the uk biobank: a mendelian randomization study . JAMA Cardiol 2017 ; 2 : 882 - 9 . 10.1001/jamacardio.2016.5804 28678979 PMC5710596 60 Voight BF Peloso GM Orho-Melander M Plasma HDL cholesterol and risk of myocardial infarction: a mendelian randomisationterm study . Lancet 2012 ; 380 : 572 - 80 . 10.1016/S0140-6736(12)60312-2 22607825 PMC3419820 61 Boyle EA Li YI Pritchard JK An expanded view of complex traits: from polygenic to omnigenic . Cell 2017 ; 169 : 1177 - 86 . 10.1016/j.cell.2017.05.038 28622505 PMC5536862 62 Merino J Leong A Posner DC Genetically driven hyperglycemia increases risk of coronary artery disease separately from type 2 diabetes . Diabetes Care 2017 ; 40 : 687 - 93 . 10.2337/dc16-2625 28298470 PMC5399655 63 Corbin LJ Richmond RC Wade KH BMI as a modifiable risk factor for type 2 diabetes: refining and understanding causal estimates using Mendelian randomization . Diabetes 2016 ; 65 : 3002 - 7 . 10.2337/db16-0418 27402723 PMC5279886 64 Ference BA Yoo W Alesh I 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 : 2631 - 9 . 10.1016/j.jacc.2012.09.017 23083789 65 Nelson MR Tipney H Painter JL The support of human genetic evidence for approved drug indications . Nat Genet 2015 ; 47 : 856 - 60 . 10.1038/ng.3314 26121088 66 HolmespersName MV Ala-Korpela M Davey SmithpersName G Mendelian randomization in cardiometabolic disease: challenges in evaluating causality . Nat Rev Cardiol 2017 ; 14 : 577 - 90 . 10.1038/nrcardio.2017.78 28569269 PMC5600813 67 Hansen LP Large sample properties of generalized method of moments estimators . Econom J Econom Soc 1982 ; 50 : 1029 - 54 . 10.2307/1912775 . Greco M FD Minelli C Sheehan NA Thompson JR Detecting pleiotropy in Mendelian randomisationterm studies with summary data and a continuous outcome . Stat Med 2015 ; 34 : 2926 - 40 . 10.1002/sim.6522 25950993 69 Ding M Huang T Bergholdt HK Nordestgaard BG Ellervik C Qi L CHARGE Consortium Dairy consumption, systolic blood pressure, and risk of hypertension: Mendelian randomization study . BMJ 2017 ; 356 : j1000 . 10.1136/bmj.j1000 28302601 PMC6168037 70 Palmer TM Nordestgaard BG Benn M Association of plasma uric acid with ischaemic heart disease and blood pressure: mendelian randomisationterm analysis of two large cohorts . BMJ 2013 ; 347 : f4262 - 4262 . 10.1136/bmj.f4262 23869090 PMC3715134 71 Afzal S Brøndum-Jacobsen P Bojesen SE Nordestgaard BG Genetically low vitamin D concentrations and increased mortality: Mendelian randomisationterm analysis in three large cohorts . BMJ 2014 ; 349 : g6330 - 6330 . 10.1136/bmj.g6330 25406188 PMC4238742 72 Wensley F Gao P Burgess S C Reactive Protein Coronary Heart Disease Genetics Collaboration (CCGC) Association between C reactive protein and coronary heart disease: mendelian randomisationterm analysis based on individual participant data . BMJ 2011 ; 342 : d548 - 548 . 10.1136/bmj.d548 21325005 PMC3039696 73 Dale CE Fatemifar G Palmer TM UCLEB Consortium; METASTROKE Consortium Causal associations of adiposity and body fat distribution with coronary heart disease, stroke subtypes, and type 2 diabetes mellitus: a Mendelian randomization analysis . Circulation 2017 ; 135 : 2373 - 88 . 10.1161/CIRCULATIONAHA.116.026560 28500271 PMC5515354 74 Sedgwick P Understanding P values . BMJ 2014 ; 349 : g4550 . 10.1136/bmj.g4550 25015369 75 Lawlor DA Tilling K Davey SmithpersName G Triangulation in aetiological epidemiology . Int J Epidemiol 2016 ; 45 :1866-86. 28108528 10.1093/ije/dyw314 PMC5841843 76 Irons DE Iacono WG Oetting WS McGue M Developmental trajectory and environmental moderation of the effect of ALDH2 polymorphism on alcohol use . Alcohol Clin Exp Res 2012 ; 36 : 1882 - 91 . 10.1111/j.1530-0277.2012.01809.x 22563891 PMC3416945 77 Mokry LE Ross S Ahmad OS Vitamin D and risk of multiple sclerosis: a Mendelian randomization study . PLoS Med 2015 ; 12 : e1001866 . 10.1371/journal.pmed.1001866 26305103 PMC4549308 78 VanderWeele TJ Tchetgen Tchetgen EJ Cornelis M Kraft P Methodological challenges in mendelian randomization . Epidemiology 2014 ; 25 : 427 - 35 . 10.1097/EDE.0000000000000081 24681576 PMC3981897 79 Swanson SA Tiemeier H Ikram MA Hernán MA Nature as a trialist? Deconstructing the analogy between Mendelian Randomization and randomized trials . Epidemiology 2017 ; 28 : 653 - 9 . 10.1097/EDE.0000000000000699 28590373 PMC5552969 80 Spitz MR Amos CI Dong Q Lin J Wu X The CHRNA5-A3 region on chromosome 15q24-25.1 is a risk factor both for nicotine dependence and for lung cancer . J Natl Cancer Inst 2008 ; 100 : 1552 - 6 . 10.1093/jnci/djn363 18957677 PMC2720751 81 Paternoster L Tilling K Davey SmithpersName G Genetic epidemiology and Mendelian randomization for informing disease therapeutics: Conceptual and methodological challenges . PLoS Genet 2017 ; 13 : e1006944 . 10.1371/journal.pgen.1006944 28981501 PMC5628782 82 Nelson CP Hamby SE Saleheen D CARDIoGRAM+C4D Consortium Genetically determined height and coronary artery disease . N Engl J Med 2015 ; 372 : 1608 - 18 . 10.1056/NEJMoa1404881 25853659 PMC4648271 83 Nüesch E Dale C Palmer TM Adult height, coronary heart disease and stroke: a multi-locus Mendelian randomization meta-analysis . Int J Epidemiol 2016 ; 45 : 1927 - 37 . 10.1093/ije/dyv074 25979724 PMC5841831 84 Davey SmithpersName G Paternoster L Relton C When will Mendelian randomization become relevant for clinical practice and public health? JAMA 2017 ; 317 : 589 - 91 . 10.1001/jama.2016.21189 28196238 85 Walker VM Davey SmithpersName G DaviespersName NM Martin RM Mendelian randomization: a novel approach for the prediction of adverse drug events and drug repurposing opportunities . Int J Epidemiol 2017 ; 46 : 2078 - 89 . 10.1093/ije/dyx207 29040597 PMC5837479 86 Fordyce CB Roe MT Ahmad T Cardiovascular drug development: is it dead or just hibernating? J Am Coll Cardiol 2015 ; 65 : 1567 - 82 . 10.1016/j.jacc.2015.03.016 25881939