@channel this is the channel for today's sessions on Mendelian Randomization.

How to copy the material for the MR practical: mkdir MR cd MR cp -r /faculty/davide/BOULDER2021/PRACTICAL1 . cp -r /faculty/davide/BOULDER2021/PRACTICAL2 .

Please also be aware - if you've downloaded ahead of time. There is a typo in Practical2.R when setting the working directory it should be:

setwd("/home/****YOURNAMEHERE****/MR/PRACTICAL2/") not PRACTICAL1. It's fixed now - so if you haven't downloaded it yet - don't worry about it!

Hi @David Evans, Given that the potential bias in MR-estimated causal effect is inversely proportional to the instrument strength (gamma), I wonder how the MR results and bias would be affected if we only used the SNPs significant at a more conservative alpha than p<5e-8; i.e., using fewer SNPs with relatively higher instrument strength vs. all the SNPs implicated in a GWAS / PGS.

1. In a PheWAS + MR study design, are we assuming that the observed phenotypic associations (in PheWAS) are necessarily due to vertical pleiotropy, before moving on to MR? Or, do we infer horizontal vs. vertical pleiotropy based on MR sensitivity analyses? Many thanks!

*Thread Reply:* Hi Madhur

1. Assuming that instrument strength is uncorrelated with size of the horizontal pleiotropic effect, as you correctly note, stronger instruments should bias estimates of the causal effect less. Stronger instruments should also be less subject to weak instrument bias (some degree of bias is always present, but its magnitude becomes less with instrument strength). Finally, the fewer instruments, the fewer opportunities for horizontal pleiotropy by definition. All of these are good things. On the other side of the coin, the more instruments in the analysis, the more power you will have, and sensitivity analyses (Egger, Modal estimator, weighted median etc) are designed to work with many rather than fewer SNPs. So, like in many areas of statistics there are advantages and disadvantages and it is a trade-off- there is never a one size fits all approach. An interesting example was an MR study of C-reactive protein I was involved in. The best strategy for an MR analysis in this study was to use a single (strongly) associated variant in the CRP gene as an instrument (i.e. pleiotropy very unlikely here and the variant had a large effect on CRP levels) rather than all of the genome-wide significant SNPs. Interestingly, these analyses yielded very different answers.
2. If by PheWAS you mean one SNP/GRS assessed against many different phenotypes, then I would say that MR and sensitivity analyses (and indeed potentially SEM approaches!) would be useful posthoc to help clarify interesting looking associations and direction of causality etc. Hope this helps! Cheers Dave

*Thread Reply:* Hi Dave, Thanks very much for the detailed answers. These are indeed helpful.

1. Your example of CRP MR was exactly what I was wondering. As a follow up, I wonder how different (or not) this issue is from the sensitivity-specificity balance one seeks to achieve in any other diagnostic/predictive biomarker.
2. In the 2. question, I was thinking of a PheWAS using the PGS of a primary phenotype (e.g. depression), but statistically, that might not be very different from using a candidate SNP (except for power). Using post-hoc SEM analyses would certainly be interesting. Many thanks, Madhur

*Thread Reply:* Yes as you say i think analogous to sensitivity-specificity concerns in diagnostics cheers Dave

For those using MR in the mental health context, there's a nice review here

Mendelian randomisation for psychiatry: how does it work, and what can it tell us?

https://www.nature.com/articles/s41380-021-01173-3