,
13 tweets,
3 min read
Read on Twitter
For the people who have not met the magic of Mendelian Randomisation (MR) it is super cool. It is rooted in epidemiology, ie, measuring things on humans to infer aspects of their health, often in the future (eg, people who smoke ... are more likely to get lung cancer)
MR is a new(ish) tool in the epidemiologists toolkit to give insight into causality - for example, people with higher BMI are more likely to get Type II diabetes. But is this *causal* - if someone reduces their weight will they be less likely to get diabetes?
You might think this is obviously yes, but how does one know? It might well be the things that make people eat more *also* independently put them at risk of Type II diabetes. This is the correlation vs causation headache of both tabloid headlines and everyday epidemiology
MR brings in genetics, not because genetics has to have a big impact, but rather using this other fact of genetics, which is that the assignment of alleles passed on from parent to child is a random choice from their parents.
This randomness is very good randomness. There are a lot of headaches about how to effectively use this randomness (how it plays out in populations not individuals being the big one) but smart geneticists can get around this/model this in a variety of ways
So - now epidemiologists have a "natural" random control trial and even better, *everyone* participates (sadly the arms of the trials are very unbalanced - most interesting alleles are less than 1% in a population, and the impact per allele is low. But... its natural trial!)
What this means is that these genetic "instruments" (borrowing terminology from econometrics which have similar statistical problems) can be used as "natural controlled experiments" and either support or refute causality.
The reality of this is that ... it works (a poster child is with lipids - MR correctly supports LDL's link to heart attacks and basically refutes HDL's link to heart attacks - the latter being a bit more complex) but ... there are a host of gotchas.
Most of these gotchas are around the need to use many (weak) instruments - it is a sort of meta-analysis-fest from other statistical areas but each meta-analysis scheme has their own assumptions
(At this point I nod wisely, try to get an emotional understanding of methods, plot beta-beta plots and be thankful the leading proponents say "run these 3 sensible meta analyses" - ideally your result should hold up in all 3)
As well as this just being cool science (we can gain big insights of the causal links between mainly environmental mediated effects using genetics. Cool or what!) it is a great example of interdisciplinary science
It needs statisticians, geneticists, epidemiologists, molecular biologists, clinicians, social scientists and pharmceutical researchers. So there is a big melting pot of ideas and perspectives.
George Davey-Smith (@mendel_random) is one of the main godfathers of this field, and he and others has attracted a great mix of people to Bristol across different disciplines - which is why the world comes to Bristol to do and think Mendelian Randomisation.
