Collaborative Targeted Maximum Likelihood Estimation to Assess Causal Effects in Observational Studies

Abstract

Observational studies can address a wide variety of issues in medicine and public health. Targeted learning (TL) provides a framework for unbiased estimation of treatment effects using these data. TL relies on two core methodologies, targeted minimum loss-based estimation (TMLE), an efficient double robust estimator, and data adaptive super learning. Collaborative TMLE (C-TMLE) is an extension of TMLE that is particularly effective when there is a sparsity of information in the data. C-TMLE provides an automated approach to constructing parsimonious models for treatment and censoring mechanisms. These models are built in response to residual bias not addressed in an initial outcome regression. C-TMLE can stabilize estimates and reduce mean squared error in high dimensional and other sparse data settings.