Abstract:
|
Instrumental variable (IV) analysis is an useful tool to address unmeasured confounding in comparing two treatment groups in observational studies. So far, the local average treatment effect (LATE) has been focused in IV analyses. In this article, we introduce a weighted local average treatment effect (WLATE), which is the LATE for a certain subpopulation of compliers, where the population of interest is determined by a function of an instrumental propensity score (PS). LATE is a special case of WLATE when the combined population of compliers is considered. We show that WLATE is expressed as a ratio of two weighted average treatment effects (WATEs). This identification result allows one to use well-developed PS methods for WATE when estimating WLATE. We have found that when one of the instrument groups is completely excluded from treatment the LATE for the eligible for the treatment is the same as the average treatment effect for the treated. Two examples of IV analyses are used to demonstrate the practical utility of the proposed methods, in which the LATEs for subpopulations are more meaningful than those for combined populations.
|