Just enter your email
When reporting estimates from multiple exposure models, such as multivariable MR (MVMR), be clear whether you are reporting direct, indirect, or total effect estimates for each exposure included in the model. This could be added to either point 6 c) or 11 b), or perhaps this is too specific for guidelines.
I think these guidelines are great. Just a couple thoughts:
I know there's only so much space in these guidelines (and in papers) but my two cents about the assumptions is that it isn't enough to mention the assumptions by name but to give a short description of what they imply in the context of the question being asked. E.g. Rather than 'we assumed the exclusion restriction' one would write, 'we assumed that the effect of the 162 SNPs in our analysis on CHD was only through their effect on education.' I think putting the assumptions in real-world terms makes it easier for non-MR people to understand what is being assumed and can even make it clearer for people who are familiar with the term 'exclusion restriction'. I advocate for this in non-MR studies as well.
I also agree with the point made by Prof Debbie Lawlor at the MR conference (if I understood correctly) about being clear that the effect should be interpreted as the effect of changing the exposure through genetic means. Although, this is captured, to a certain extent, by 16 c) which asks whether other interventions could have the same size effect.
Also, just a thought, are we at the point now where the guidelines should not just be, "State whether statistical code is publicly accessible and if so, where," but instead, "Make statistical code publicly accessible and indicate where." Of course, there may be special circumstances where it's not possible to share code, but otherwise, it is very easy for anyone to share their code now and it's an important part of transparency and reproducibility.
Thanks! Great work!
This looks great.
I have a few suggestions:
1. Perhaps you could add additional guidelines on how to report "null results". For example, studies often make exaggerated claims that "there was no evidence for an effect", without considering whether their estimated effects were compatible with expected effect sizes (e.g. from observational studies), taking into account confidence intervals.
2. In general, studies should be encouraged to consider a priori or expected effect sizes (e.g. from observational studies), which can help guide interpretation of MR results, e.g. statistical power as well as interpretation of null results.
3. Explicit consideration of statistical power and instrument strength.
4. The list seems largely designed with one sample MR analyses in mind and there are a lot of items that aren't really relevant or practical for two sample MR studies. For example, many of the items in section "4. Study design and data sources".
5. Some of the suggested items will be extremely difficult or impractical for two sample MR studies. For example, "For each exposure, outcome and other relevant variables, describe methods of assessment and, in the case of diseases, the diagnostic criteria used." Sometimes a two-sample MR study will be based on very large meta-analyses of exposure and/or outcome studies, where going back to the original studies to identify all this information is likely to be practically impossible. I agree that just because something is difficult doesn't mean we shouldn't do it. But there might be alternative metrics that are easier to get hold of that also address the same underlying issues, e.g. metrics of heterogeneity between studies within the outcome or exposure GWAS meta-analysis.
6. The recommendation in 10.d.i. will usually be quite difficult to obtain. Instead, authors should be encouraged to provide evidence that the exposure and outcome samples come from the same population, which could include 10.d.i on rare occasions. "Provide information on the similarity of the genetic variant-exposure associations between the exposure and outcome samples."