Conference Agenda

The predictive power of machine learning is often celebrated, but caution is also warranted due to the potential for algorithmic bias which often arises from classical statistical concerns such as confounding and selection bias. Statisticians are thus often wary of the use of machine learning in the context of treatment recommendations and other highly sensitive and potentially life-altering decision-making. I will discuss two examples in which machine learning approaches were incorporated into a classical statistical method to learn individualized treatment strategies that are designed to address confounding to yield causally valid conclusions. In the first, non-parametric ensemble learner is used in the context of an approach that is not robust to model mis-specification. In the second, probabilistic supervised learning in the form of Gaussian processes will be used to improve performance of inverse probability of treatment weighted estimators. In both cases, the use of machine learning is layered onto classical statistical approaches to causal inference that have been developed to address confounding in the context of observational data analysis. Relevant causal assumptions, and how they may (or may not!) be detected and mitigated will also be discussed.