Re: "Credible Mendelian Randomization Studies: Approaches For Evaluating The Instrumental Variable Assumptions"

The recent article by Glymour et al. (1) on evaluating instrumental variable (IV) assumptions in Mendelian randomization contains some useful guidance for empirical researchers for detecting violations in assumptions which may lead to seriously biased estimates of causal associations. Consider the 4 equivalent statements made in the section entitled “Falsifying IV Assumptions by Leveraging Prior Causal Assumptions.” Assuming that the direction of confounding is positive (with equivalent statements if the confounding is negative), the simplest of these statements is that the IV estimate should be less than the ordinary least squares estimate for the observational regression of outcome on exposure. Other than in extreme nonlinear scenarios, the negation of this statement is claimed as a decisive indication that the IV is not valid. Although this statement is true for the underlying parameters, it is not generally true for the estimates. We can show this in a simple simulation exercise. Additionally, we may not expect to obtain the same answer from a Mendelian randomization approach as from a conventional epidemiologic analysis, so this statement may not be a fair assessment of the validity of the IV.