Disentangle the Semiparametric Structure in Statistical Models, with Applications in Causal Inference and Model Transportability

This short course offers statisticians and data scientists a comprehensive overview to disentangling semiparametric structures in statistical models, with applications in missing data analysis, causal inference, high-dimensional genetics data, and evaluating model transportability and generalizability. Participants will explore both classical and cutting-edge semiparametric techniques, with emphasis on their importance in achieving efficiency, robustness, and their tradeoff. Statistical inference often relies on models with assumptions about the data-generating process, which sometimes are parametric, defined by a finite-dimensional parameter. However, parametric models have limited robustness. Semiparametric models, blending parametric and nonparametric nuisance components, offer a flexible framework to capture complex relationships without imposing overly restrictive assumptions. Semiparametric structures have become critical in model-agnostic frameworks, and have led to emerging applications in causal inference, precision medicine, domain adaptation, and model generalization. This short course will be divided into two parts: Part 1 (2 hours): A review of fundamental concepts in parametric models, focusing on semiparametric structures within them, followed by an introduction to extending these ideas into semiparametric models and the derivation of the efficient influence function. Part 2 (2 hours): Two real-world applications: estimating treatment effects in causal inference, and evaluating model transportability from randomized controlled trials to a target population.