As compared with standard IV estimators, subsample IV estimators exhibit different behavior in the presence of weak IVs. For traditional IV estimators, estimates are biased towards the confounded observational association. In contrast, subsample IV estimators are biased towards the null when the subsample is relatively small, similar to 2-sample IV estimators (5). However, they are biased towards the confounded association when the subsample is increased, similar to traditional IV estimators.
