We introduce the Noisy-Or Risk Allocation (NORA) model, for generating new causal knowledge, which considers multiple independent causes in a setting with binary features and outcomes. NORA infers the risk of each exposure and we hypothesize that it is more robust to confounding by backdoor paths. The proposed methods may permit researchers across domains and institutions to optimize outcomes with interventions, without the need for a trial.
Learning Objective 1: After participating in this session, the learner should be better able to:
- Understand the methodological weaknesses of current, popular methods of causal inference from observational data
- Articulate the principals of a new machine learning method for causal inference, NORA, and describe the model’s strengths and weaknesses.
- Describe potential-use cases for the proposed model, NORA.
Amelia Averitt (Presenter)
Adler Perotte, Columbia University