Adapting black-box machine learning methods for causal inference
I’ll discuss the use of observational data to estimate the causal effect of a treatment on an outcome. This task is complicated by the presence of ‘confounders’ that influence both treatment and outcome, inducing observed associations that are not causal. Causal estimation is achieved by adjusting for this confounding by using observed covariate information. I’ll discuss the case where we observe covariates that carry sufficient information for the adjustment, but where explicit models relating treatment, outcome, covariates, and confounding are not available. For example, in medical data the covariates might consist of a large number of convenience health measurements of which only an unknown subset are relevant, and even then in some totally unknown manner. Or, the covariates might be a passage of (natural language) text that describes the relevant information. I’ll describe how to modify standard architectures and training objectives to achieve statistically efficient and practically useful causal estimates, as well as how to adapt traditional approaches to evaluating sensitivity to unobserved confounding to allow for the use of blackbox models.
Victor Veitch is a distinguished postdoctoral research scientist in the department of statistics at Columbia University. He completed his PhD in Statistics at the University of Toronto. His work addresses both the use of machine learning for causal inference, and the modeling of relational and network data. He has been recognized with a number of awards, including the 2017 Pierre Robillard award for best Statistics PhD thesis in Canada.