Causal Inference in Randomized Trials with Noncompliance and Losses to Follow-Up

• Activity Number: 488
• Type: Topic Contributed
• Date/Time: Wednesday, August 6, 2014 : 10:30 AM to 12:20 PM
• Sponsor: Biometrics Section
• Author(s): Miguel Hernan*+
• Companies: Harvard School of Public Health
• Keywords: causal inference ; randomized trials

Abstract

An intention-to-treat analysis is commonly used in randomized trials to estimate the effect of assignment to treatment. However, intention-to-treat effects are problematic in randomized trials with substantial noncompliance or losses to follow-up. These deviations from protocol, which often arise in trials that study the long-term effects of sustained clinical interventions in typical patients and care settings, may induce postrandomization confounding and selection bias. These two biases need to be adjusted for when one is interested in the treatment effect if patients had adhered to the study protocol. i.e. , the per protocol effect. There are two main approaches to consistently estimate this per protocol effect: 1) to analyze the randomized trial as if it were an observational study, i.e., to adjust for measured time-varying covariates while assuming that no important covariates were unmeasured, and 2) to exploit the randomization indicator as an instrumental variable in order to adjust for bias without having to measure any covariates. This talk presents novel applications of each of these two approaches to randomized trial data and discusses the relative advantages and disadva