Targeted Learning

Topics

• Query subject-matter experts, study designers, and others involved in the data generation process in order to ○ use the data at-hand to formulate a statistical estimation problem that’s realistic, reliable, and reproducible; ○ check assumptions that are required to use this data to answer causal questions (i.e., identifiability); ○ build a library of machine learning algorithms that is consistent with the processthat generated the data; ○ design improved (potentially hypothetical) studies that would permit answering the relevant scientific question of interest, and evaluate if this ideal experiment is possible/ethical in the real-world; and ○ translate the final scientific question of interest into a statistical question whose answer can be reliably estimated from the data. Design an ensemble of machine learning algorithms, Super Learner, such that the optimality theory (i.e., oracle inequalities for the general cross-validation selector) is likely to hold for the data at-hand. Explain the importance of asymptotic linearity and efficiency/inference to non-statistician collaborators, and reason whether or not the current estimator meets these conditions. Use R’s tlverse software ecosystem to ○ define a machine learning task that mirrors your estimation problem; ○ specify the question of interest in terms of a tlverse Spec; and ○ design simulations that can be used to evaluate the behavior of estimators and their inference, ultimately informing your final choice for estimation.

Official catalogue course objectives and student learning outcomes