Figure 2.

Causal structures that were assessed in the simulation study.$Z$: genetic instrument; $W$: covariable; $R$: possible direct consequence of $Z$; $X$: exposure; $Y$: outcome: $U$: unmeasured common cause. Rows represent different causal structures between $Z$, $X$, $W$ and $Y$ as illustrated by Scenarios A ($W$ fully mediates the effect of $Z$ on $X$), B ($Z$ independently affects both $X$ and $W$), C ($X$ fully mediates the effect of $Z$ on $W$), D (effect of $Z$ on $X$ and $W$ is mediated by a common cause $R$), E (same as B except that $W$ has a direct effect on $Y$) and F (same as D except that $W$ has a direct effect on $Y$). Columns represent different confounding structures between $X$, $W$ and $Y$ as illustrated by Scenarios A1-F1 (no unmeasured confounders other than R), A2-F2 (presence of $X$-$W$ confounder: $U_{X,W}$), A3-F3 (presence of $W$-$Y$ confounder: $U_{W,Y}$), A4-F4 (presence of $X$-$Y$ confounder $U_{X,Y}$) and A5-F5 (presence of all three confounders simultaneously)