A0894
Title: Incorporating mechanistic knowledge in causal inference
Authors: Alexander Volfovsky - Duke University (United States) [presenting]
Abstract: At their core, the assumptions needed for causal inference are concerned with removing the effects of potentially unobserved quantities. We may know that a drug is given, but maybe not when, or we may observe where a disease is transmitted but maybe not exactly from whom, yet in both settings, we might be interested in causal questions: Does the drug have an effect? Does a mitigation strategy work to prevent future transmission? Because these processes are governed by established biological mechanisms, mechanistic models can provide invaluable insights into the interactions between biological objects (drug diffusion in the body, transmission probabilities between individuals). Conditioning on these models can provide more credibility to the necessary assumptions for causal inference. We present two case studies of leveraging these types of models for causal inference: (1) we analyze observational data of critically ill patients and identify the effect of seizures if they were not treated, and (2) we employ a mechanistic model of disease transmission to help design a trial for evaluating a non-pharmaceutical intervention.