User:Lvdlaan/sandbox

What is it?
Targeted Minimum Loss Estimation (TMLE) is a general statistical framework for developing semi/nonparametric efficient substitution estimators for statistical quantities of interest. Notably, TMLE utilizes Machine-Learning (ML) to flexibly estimate relevant nuisance parameters of the data-generating distribution; thereby allowing for valid statistical inference with minimal assumptions. Targeted learning has led to significant theoretical development in the statistical fields of causal inference, survival analysis , variable importance , analysis of longitudinal data , network   and time-series analysis , and contextual bandit  problems. Targeted learning has also found extensive applications in clinical trials, observational studies, epidemiology, safety analysis, and personalized healthcare. TMLE was developed by U.C. Berkeley Professor of Statistics and Biostatistics, Mark J. van der Laan, and first introduced in the seminal paper "Targeted Maximum Likelihood Learning" in 2006. Since then, TMLE has become an active field of research in both statistical theory and application. As of July 2021, hundreds of peer-reviewed articles and two books have been published on the topic.

TMLE can be viewed as a two-step procedure. In the first step, one obtains an initial estimator of data-generating distribution (or relevant nuisance parameters thereof), possibly utilizing black-box machine learning algorithms. In the second step, the so-called targeting step, one de-biases the initial estimator for the target parameter of interest by performing minimum loss estimation along a clever parametric perturbation sub-model.

utilizing machine-learning.

TMLE has found extensive applications in causal inference, randomized and observational studies

The TMLE template requires first defining the statistical model -- a collection of probability distributions for the observed data that contains the true data-generating distribution. Given the statistical model, one can then define

TMLE is related to the efficient one-step and estimating equation   methodologies in that it utilizes the efficient influence function of the target parameter of interest to debias an initial.