Discover and read the best of Twitter Threads about #ivestimation
Most recents (2)
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...
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...
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.
This is the routine approach in RCTs to test the null. It’s called intent-to-treat approach.
Anyone who says “Z is an instrumental variable” doesn’t know what an instrumental variable is.
We can only say “Z is a proposed instrumental variable”, because we cannot ever verify that Z is an instrumental variable.
Not even in many randomized #trials.
journals.lww.com/epidem/fulltex…
We can only say “Z is a proposed instrumental variable”, because we cannot ever verify that Z is an instrumental variable.
Not even in many randomized #trials.
journals.lww.com/epidem/fulltex…
Of the 3 instrumental conditions an IV must meet
(1) is empirically verifiable
(2) isn't verifiable, but expected in double-blind trials
(3) isn't verifiable, but expected in randomized trials
Thus #IVestimation will be biased in many #trials because condition (2) doesn't hold.
(1) is empirically verifiable
(2) isn't verifiable, but expected in double-blind trials
(3) isn't verifiable, but expected in randomized trials
Thus #IVestimation will be biased in many #trials because condition (2) doesn't hold.
Though we can't verify that a variable is an IV, we can sometimes falsify that a variable is an IV. @MariaGlymour et al explain how
academic.oup.com/aje/article/17…
One would then expect all IV (eg, Mendelian randomization) studies try to falsify but, amazingly, most don’t. Malpractice.
academic.oup.com/aje/article/17…
One would then expect all IV (eg, Mendelian randomization) studies try to falsify but, amazingly, most don’t. Malpractice.
