Methods to Adjust for Treatment Switching That Disrupts Randomization in Clinical Trials

• Activity Number: 488
• Type: Topic Contributed
• Date/Time: Wednesday, August 6, 2014 : 10:30 AM to 12:20 PM
• Sponsor: Biometrics Section
• Author(s): Jessica Young*+
• Companies: Harvard School of Public Health
• Keywords: causal inference ; treatment switching ; g-formula ; inverse probability weighting

Abstract

Causal inference based on randomized clinical trials may be compromised when some subjects fail to follow their assigned treatment strategy at some time after the randomization. For example, a patient assigned to placebo might be switched to an active treatment if she becomes very ill. Alternatively, a patient on active treatment might be switched after experiencing severe side effects. In this talk, we review Robins' g-formula (1986) as a means of identifying and estimating the causal effect of following different treatment strategies in a randomized trial where treatment switching occurs. Specifically, we review sufficient identifying conditions in a general time-varying setting where (i) measured patient characteristics at a given time are associated with both future outcomes and future treatment and (ii) these characteristics are themselves affected by past treatment. We also review two approaches to effect estimation motivated by these identification results which may be applied in typical high-dimensional settings. These include the parametric g-formula and inverse probability weighted estimation.