Today Mendelian Randomization (MR) is usually implemented as a form of instrumental variable (IV) estimation.
Aware that valid #IVestimation requires strong assumptions, MR advocates often retreat to the position that MR numerical estimates need not be taken seriously...
Their position: The goal of MR is to "test causality". MR studies aren't designed to yield a valid numerical IV estimate of causal effect. MR studies are designed to answer a yes/no question: Is the causal null hypothesis true?
But retreating to *null testing* is problematic...
1) If the goal of MR is simply to test the null, then why use #IVestimation at all? MR papers should just report the association between the genetic trait (proposed IV) and the outcome.
This is the routine approach in RCTs to test the null. It’s called intent-to-treat approach.
2) If the goal of MR is simply to test the null, we don't need #IVestimation but we still need an IV.
That is, all criticisms of genetic traits as IVs continue to apply even if the goal is redefined as "just testing the null." See recent discussion here
3) "Retreating to null testing" raises other serious problems when, as in most MR studies, the underlying exposure is time-varying. Even if the genetic trait were a true IV!
Sonja Swanson, Jeremy Labrecque (@radarlake) and I discuss IV null testing here:
link.springer.com/article/10.100…
Share this Scrolly Tale with your friends.
A Scrolly Tale is a new way to read Twitter threads with a more visually immersive experience.
Discover more beautiful Scrolly Tales like this.