Orthogonal Statistical Learning with Self-Concordant Loss

As statistical machine learning impacts several domain applications of major importance to the planet and society, ranging from healthcare to the environment, sophisticated approaches to estimation, proceeding in multiple stages, are being developed to overcome confounding factors and to address high-dimensional nuisance parameters (Peters et al., 2017). Orthogonal statistical learning (OSL), and its statistical estimation predecessor double machine learning (DML), have emerged as general frameworks for two-stage statistical machine learning in the presence of a nuisance component (Mackey et al., 2018; Liu et al., 2021; Nekipelov et al., 2022). The power of this framework can be illustrated on the task of assessing the causal effect of a treatment on an outcome of interest.