Causal inference (CI) is a methodology to uncover causal relationships in real-world problems. Given the explosion in the volume and types of data available to support outcomes research, health economics, and epidemiology, it has been argued that deep/machine (DL/ML) learning may also provide renewed vitality to causal inference, with new ideas in CI able to promote new advancements in ML/DL. Machine and deep learning (DL) for causal inference is still evolving – e.g., in natural language processing recently, researchers have aimed to enhance natural language processing tasks with causal inference.
Predictive analytics are rapidly being upgraded using machine learning methods – however a question remains as to how to draw causal inference from observational data such as electronic health records. Can machine learning help? Minimally ML would be an effective for hypothesis generation, as ML's strength is in identifying correlational structures in observational data – these structures can then be tested with usual causal modelling approaches.
Can we use machine learning (deep learning) to estimate causal models directly? Can we incorporate a statistical perspective into data science? These questions are especially relevant in applications to stochastic processes (COVID-19, diseases) and big data such as electronic health records (eHRs) and clinical care, precision medicine, public and global health, digital health, and even now in randomised controlled trials (RCTs).
This session provides a place to explore such questions and the connections between classical statistical methods and machine learning, deep learning models and causal inference. Topics relating to tools used for intervention tasks, ML/DL, Counterfactuals causal mediation, Bayesian networks, Random Forests, Neural networks, and inverse-probability weighting for time-varying treatments and targeted learning leveraging machine learning algorithms are welcome.
Key topics: Data science tools, with a focus on causal inference, Tools used for intervention tasks, Explainability versus causality in eHRs, Propensity scores