École Nationale Supérieure de
Techniques Avancées

Abstract

Causal inference is a research field in statistics that aims to understand the relationships between an exposure (or treatment) and an outcome. This field is proficient in various domains such as social sciences, healthcare, and economics. Mediation analysis, a component of causal inference, aims to identify and quantify the indirect effect of the exposure on the outcome through one or more intermediate variables called mediator(s). There, conducting estimation implies a correct study design, meticulous mediator selection, and thorough consideration of confounding factors. This internship takes part during the writing of the article authored by Judith Abecassis, Julie Josse, and Bertrand Thirion (2022), and find its contribution in the development of the med_bench package within the SODA team at INRIA Saclay lab. The med_bench package is intended to enrich the accessibility of various mediation estimators such as coefficient product, g-computation, and inverse probability weighting (IPW). These estimators, some limited to R or Python, will be harmoniously implemented with different variations and their performances will be compared. The contributions outlined in this report lie in the package enhancement, through a rigorous documentation and a range of tests built using pytest, as well as two new Python implementations for causal mediation analysis, the multiply-robust estimator and the double machine learning estimator (DML). We also provide an evaluation of the performances of these two implementations.