We assessed weak IV bias for subsample IV scenarios varying the R2 for the regression of X on G (rather than F), as R2 may be a more meaningful parameter to MR practitioners (Figure 4). However, because weak IV bias is related to F rather than to R2, bias does not vary with the nX:nY ratio in a uniform way, since increasing nX both increases the proportion of participants in the subsample (leading to a greater bias towards the confounded association) and increases the F statistic, leading to a reduction in weak IV bias. Thus, weak IV bias towards the confounded association is much more pronounced when nY is small, because nY limits the size of nX, reducing the F statistic.
