Using Knowledge of the Data Structure in Applications of the Parametric G-Formula

• Activity Number: 233
• Type: Topic Contributed
• Date/Time: Monday, August 10, 2015 : 2:00 PM to 3:50 PM
• Sponsor: International Chinese Statistical Association
• Author(s): Jessica Young*
• Companies: Harvard School of Public Health
• Keywords: g-formula ; treatment switching ; causal inference ; randomized trials

Abstract

Given sufficient time-varying measurements of patient characteristics are available, Robins' g-formula may identify the causal effects of time-varying treatments in the presence of time-varying selection bias.  This bias may occur in randomized trials when subjects fail to comply with their baseline assigned treatment strategy during the follow-up.  Estimating causal effects via the g-formula is challenged by the curse of dimensionality.  Several methods have been proposed to deal with this problem including the parametric g-formula, inverse probability weighting and doubly robust methods.  In this talk, we review the parametric g-formula approach to estimating this high-dimensional function of the data.  This approach has the advantage of relying on off-the-shelf subroutines available in all statistical software.  However, model misspecification is a particular concern with this method and can be guaranteed to some degree under the causal null.  Here we consider how reliance on arbitrary parametric assumptions may be minimized with this approach by incorporating all available knowledge of the data structure.  We give examples from previous applications in observational studies.