medoutcon: Nonparametric efficient causal mediation analysis with machine learning in R

Abstract

Science is most often concerned with questions of mechanism. In myriad applications, only the portion of the causal effect of an exposure on an outcome through a particular pathway under study is of interest. The study of such path-specific, or mediation, effects has a rich history, first undertaken scientifically by Wright (1921) and Wright (1934). Today, the study of such effects has attracted a great deal of attention in statistics and causal inference, inspired by applications in disciplines ranging from epidemiology and vaccinology to psychology and economics. Examples include understanding the biological mechanisms by which vaccines causally alter infection risk (Benkeser et al., 2021; Hejazi et al., 2020), assessing the effect of novel pharmacological therapies on substance abuse disorder relapse (Hejazi et al., 2022; Rudolph et al., 2020), and evaluating the effects of housing vouchers on adolescent development (Rudolph et al., 2021). The medoutcon R package provides researchers in each of these disciplines, and in others, with the tools necessary to implement statistically efficient estimators of the interventional direct and indirect effects (Dıáz et al., 2020) (for brevity, henceforth, (in)direct effects), a recently formulated set of causal effects robust to the presence of confounding of the mediator-outcome relationship by the exposure. In cases where such confounding is a nonissue, the interventional (in)direct effects (VanderWeele et al., 2014) reduce to the well-studied natural (in)direct effects (Pearl, 2001; Robins & Greenland, 1992), for which medoutcon provides efficient estimators similar to those of Zheng & van der Laan (2012). By readily incorporating the use of machine learning in the estimation of nuisance parameters (through integration with the sl3 R package (Coyle, Hejazi, Malenica, Phillips, & Sofrygin, 2021) of the tlverse ecosystem (van der Laan et al., 2022)), medoutcon incorporates state-of-the-art non/semi-parametric estimation techniques, facilitating their adoption in a vast array of settings. Statement of Need While there is demonstrable interest in causal mediation analysis in a large variety of disciplines , thoughtfully implementing data analysis strategies based on recent developments in this area is challenging. Contributions in the causal inference and statistics literature largely fall into two key areas. Broadly, the study of identification outlines novel causal effect parameters with properties desirable in real-world settings (e.g., the interventional effects, which can be learned under mediator-outcome confounding) and untestable assumptions under which a statistical functional corresponds to a causal estimand of interest. A complementary line of study develops non/semi-parametric efficiency theory for the statistical functionals outlined in the causal identification literature, allowing for their robust estimation with modern techniques from machine learning. Neither concerns itself with opening the door to applying these estimators in real-world data analyses. Moreover, the implementation of open source software Hejazi et al., (2022). medoutcon: Nonparametric efficient causal mediation analysis with machine learning in R. Journal of Open Source Software, 7(69), 3979. https://doi.