Abstract
The Targeted Maximum Likelihood Estimation, proposed in 2006, is a general framework for developing efficient asymptotically linear estimators for general parameters in large (semi-parametric) statistical models. It originally provided an alternative to estimation equation (EE) approaches (such as inverse probability of treatment weighted, IPTW, and augmented IPTW estimators) with some finite sample benefits and provided estimation for parameters/models where EE estimators are not available or do not lead to single solutions. TMLE relies on initial fits of the relevant components of the data-generating distribution, and the emphasis has been on machine learners with explicit optimality properties (i.e., the SuperLearner). One clear advantage of TMLE is its likelihood-based model selection, which has led to ad hoc but very powerful finite-sample augmentations that lead to more robust “machine-like” performance, such as collaborative TMLE (C-TMLE). Another important augmentation is cross-validated TMLE (CV-TMLE), which provides additional protection against overfitting and, relatedly, fewer constraints on the underlying model and fitting algorithms (i.e., Donsker class constraints). Most of the specific algorithms available in the software (mainly R) are for estimating specific estimands inspired by causal inference (e.g., the SuperLearner/tmle packages on CRAN, the sl3/tmle3 packages on Git Hub). Now, a wide variety of algorithms are available for many more complex parameters (longitudinal treatment effects, mediation effects, optimal treatment rules, etc.). TMLE and the general Targeted Learning framework is an active area of research, with new improvements to both the asymptotic and finite sample behavior being proposed regularly. In this talk, I will provide a general background and highlight some of these augmentations and future directions.