Abstract
Mendelian randomization, i.e., instrumental variable analysis with genetic instruments, is an increasingly popular and influential analytic technique that can foreshadow findings from randomized controlled trials quickly and cheaply even when no study measuring both exposure and outcome exists. Mendelian randomization studies can occasionally generate paradoxical findings that differ from estimates obtained from randomized controls trials, potentially for important, yet to be discovered, etiological reasons. However, bias is always a possibility. Here we demonstrate, with directed acyclic graphs and real-life examples, that Mendelian randomization estimates may be open to quite severe bias because of deaths occurring between randomization (at conception) and recruitment years later. Deaths from the genetically predicted exposure (survival bias) and when such deaths occur deaths from other causes of the outcome (competing risk) both contribute. Using a graphical definition of the condition for a valid instrument as “every unblocked path connecting instrument and outcome must contain an arrow pointing into the exposure” draws attention to this bias in Mendelian randomization studies as a violation of the instrumental variable assumptions. Mendelian randomization studies likely have the greatest validity if the genetically predicted exposure does not cause death or when it does cause death the participants are recruited before many such deaths have occurred and before many deaths have occurred due to other causes of the outcome.
Footnotes
Conflicts of interest: There is no conflict of interest
Sources of financial support: None
Data & Code: The data used here is publicly available, the code is available on request.