fixed for each ratio. Figure 3 (right panel) shows the same simulations conducted in the absence of confounding; hence, no bias towards the confounded association is observed. For weak IVs, bias towards the null is present for all subsample scenarios, and the estimate moves closer to the null as nX:nY decreases. For the 2-sample approach (Web Figure 3), bias towards the null increases as F decreases, regardless of the value of nX:nY. This is true for both the confounded and unconfounded scenarios.
