Dealing with Time-Varying Eligibility for Exposure Using the Target Trials Approach to Causal Inference with Electronic Health Records

• Activity Number: 543 - Making Sense of Complex Featured Data with Statistical Methods
• Type: Invited
• Date/Time: Wednesday, July 31, 2019 : 2:00 PM to 3:50 PM
• Sponsor: SSC
• Author(s): Mireille Schnitzer*
• Companies: University of Montreal
• Keywords: causal inference; target trial; longitudinal; marginal structural models; health application; perinatal

Abstract

Improper handling of time-dependent exposures can result in ill-defined causal effects, residual time-dependent confounding, and immortal time bias. Marginal structural models were proposed as a way to define effects when exposures are time-varying with confounders affected by prior treatment. However, typical definitions of exposure effects no longer apply when eligibility for exposure varies over time. We demonstrate how the concept of “treatment strategies” in the context of a target trial (Hernán and Robins, 2016) provides a solution in such situations using two real examples. The first example involves post-baseline contraindications to Direct Oral Anticoagulants (DOACs) in a study that aims to contrast DOACs with warfarin in patients with atrial fibrillation. The second example involves contrasting trimester-specific exposures during pregnancy when some women deliver in the second trimester and the time of delivery may be affected by earlier exposure. In both of these examples, we define effects based on potential outcomes under relevant treatment strategies and show how these effects can be estimated with inverse probability of treatment weighting.