In their recent @ajhgnews.bsky.social Review, @vkarhune.bsky.social, @stevesphd.bsky.social, & co discuss the key considerations and provide advice to produce a higher standard in planning, conducting, reviewing, and interpreting cis-Mendelian randomization studies: https://bit.ly/48DEz0H #ASHG
Thanks Kaur! Your comments were really useful and helped ensure that we got the message across clearly. Much appreciated!
Thanks to all co-authors, and also to reviewers who helped and supported this work (including @kauralasoo.bsky.social , who kindly self-identified!) - comments and feedback welcome!
Cis-MR studies are not intrinsically superior to genome-wide MR studies, and algorithmically-performed cis-MR analyses will rarely be optimal. But when performed with care, cis-MR is a powerful tool to inform about putative causal effects.
Validation can be statistical (e.g. colocalization), but also should be biologically-motivated (e.g. positive and negative controls). Robust MR methods are rarely conclusive, as if one variant is invalid, it is likely nearby variants will be invalid in a similar way.
In particular, molecular exposure biomarkers measured in bulk tissues (usually blood) may reflect mechanisms in irrelevant tissues, whereas downstream biomarkers may be more specific.
The exposure biomarker reflects the causal mechanism of interest. It is not necessarily the causal risk factor itself, and often choosing a downstream trait is better - if water has reached a proximal downstream station, then it must have passed though the causal mechanism.
Important steps are: 1) defining the analysis question, 2) choosing an appropriate gene region, 3) choosing an appropriate biomarker of the causal mechanism, 4) choosing the optimal variant(s) in the selected gene region(s), 5) validating the variants.
Our review provides guidance on how to perform reliable cis-MR analyses - the analysis is often not difficult to implement, and the major effort is ensuring that you are performing the optimal analysis, and reporting that analysis in a reasonable way.
It's tempting to this of cis-Mendelian randomization analyses, particularly those using a single genetic variant, as simplistic, and analysis with multiple variants using large numbers of methods as more sophisticated. However, "design trumps analysis"...
Our review "Integrating genetic data with biological insight: A practical guide to cis-Mendelian randomization" is now published at @ajhgnews.bsky.social - led by @vkarhune.bsky.social and Benji Woolf with critical insight from Dipender Gill and Pallav Bhatnagar. Thread follows:
Thanks to Ash for leading this work, and to Frank DiTraglia for asking difficult questions about the statistical methodology. All comments (and suggestions for applications) welcome!
Lots of interesting maths under the hood here in terms of flexibly estimating the function defining individuals' stickiness to change their behaviour (the propensity score), and ensuring our estimates are robust and efficient to the specification of this function.
Why? Many potential reasons, but my leading theory is that these genetic variants lead to a greater change amongst drinkers not just in drinking excessively, but drinking hazardously (e.g. binge drinking), hence a greater effect on blood pressure.
In UK Biobank, we observed the opposite. The genetic effect predisposing individuals to greater average alcohol consumption is associated with a greater increase in blood pressure in those who choose to drink than those who choose not to drink - reverse selection on gains.
We investigate heterogeneity in the effect of excessive alcohol consumption on blood pressure using genetic variants as instruments (this is the 'policy'). Is the harmful effect weaker in those who currently choose to drink? (Here, high blood pressure is bad, so rationally we expect ATT<ATU.)
If the ATT is greater than the ATU, then people who choose take the treatment are those who benefit most from the treatment: this is known as "selection on gains". If there is treatment effect heterogeneity, then we may rationally expect to see selection on gains.
If we are interested in the effect of a policy that offers the treatment to those who currently take the treatment, then the ATT is important; for the effect on those who do not currently take the treatment, then the ATU is important.
Other times, the target population is those with the exposure (average treatment effect in the treated - ATT), or those without the exposure (average treatment effect in the untreated - ATU). This allows us investigate if there is treatment effect heterogeneity.
Any epidemiological estimate represents a quantity estimated in a specific dataset, and targets a quantity defined for a specific population. Sometimes this is the whole population, as for the average treatment effect (ATE).
New pre-print on treatment effect heterogeneity led by Ash Patel: "Efficient semiparametric estimation of marginal treatment effects with genetic instrumental variables" available at arxiv.org/abs/2603.08871. Brief summary:
Non-linear MR did not show evidence for non-linearity for most outcomes. Where it did, it never suggested a non-monotone relationship - no J-shaped or U-shaped findings for any outcomes.
We threw MR with alcohol as an exposure at a large number of exposures. Most came out supporting harmful effects, particularly for neurologic and behavioural, circulatory, and liver outcomes. Potential protective effects were for migraines and urinary calculus.
Great to be involved in this publication: "Phenome-wide study on alcohol consumption provides genetic evidence for a causal association with multiple diseases and biomarkers" led by Nigussie Kassaw - doi.org/10.1016/j.nu....
Keir Starmer isn't a person? (The quote isn't specifically relevant to the topic, but he is quoted in the piece!)
Great to see the first paper from James' PhD out as a pre-print - thanks to Ash for providing lots of support and help! Look forward to receiving comments!
Using MSE, with lots of confounding (\rho>0.3), IV outperforms OLS at F lower than 10. With minimal confounding (\rho~0.1), the F threshold is higher. Using an F statistic to determine your analysis strategy is a bad idea in any case, but that's a story for another day.
Relative bias is not the best criterion to focus on - it gives a simple rule, but not a helpful rule. Mean squared error or mean absolute error give more helpful rules, but these rules will depend (correctly!) on the degree of confounding between the exposure and outcome.
The F>10 threshold is independent of the amount of confounding between the exposure and outcome. But the choice between IV and OLS *should* be dependent on this - if there is lots of confounding, we would typically prefer IV. If no confounding, we prefer OLS.
This is particularly due in the asymptotic limit under standard weak instrument asymptotics, where the OLS estimate tends to a point, but the IV estimate tends to a distribution. Relative bias only cares about the behaviour of the average of this distribution.