In this report, we compared four estimators (intent-to-treat, as treated, per protocol, and instrumental variables estimators) that are conventionally considered for treatment effect estimation by simulation under different non-compliance scenarios in typical clinical trial settings. We found that intent-to-treat and instrumental variables estimators are not perfect and can be problematic in some situations although these two estimators carry desirable properties as we assume.
Copyright (c) 2006 John Wiley & Sons, Ltd.