Collaborative Targeted Maximum Likelihood Estimator (CTMLE) in Observational Studies

• Activity Number: 227
• Type: Invited
• Date/Time: Monday, August 10, 2015 : 2:00 PM to 3:50 PM
• Sponsor: Health Policy Statistics Section
• Author(s): Mark Johannes van der Laan* and Sam Lendle and Sebastian Schneeweiss and Cheng Ju
• Companies: UC Berkeley and UC Berkeley and Brigham and Women's Hospital and UC Berkeley
• Keywords: Causal inference ; Targeted Maximum Likelihood Estimation

Abstract

We consider the problem of estimation of a marginal causal effect of a single time-point binary treatment on a binary rare outcome in observational studies. TMLE is an efficient substitution estimator relying on a fit of the propensity score to update an initial estimator of the regression of outcome on treatment and covariates. One of the challenges in the construction of TMLE and other estimators proposed in the literature is how to fit the propensity score/treatment mechanism so that it actually results in the desired bias reduction, mitigating the risk of including covariates that are only predictive of treatment and the risk of excluding actual important confounders, in the context that one has to search among a very large collection of potential confounders. Collaborative TMLE (e.g, van der Laan, Gruber, 2008) is an effective template to build estimators of the propensity score, since it iteratively selects variables into the propensity score model for which the TMLE update of the outcome regression is maximally significant. In this talk we review C-TMLE, consider implementations that are scalable to large dat sets, and evaluate performance based on real and simulated data.