Primary Submission Category: Generalizability/Transportability
Harmonizing graphical methods for dynamic and static causal modelling
Authors: Ian Shrier, Naftali Weinberger, Tyrel Stokes, Russell J. Steele,
Presenting Author: Ian Shrier*
Causal directed acyclic graphs (DAGs) succinctly outline assumptions about causal relationships between variables. Causal DAGs are called static models because the random variables are in a stable state. A light switch either completes or interrupts an electrical circuit. However, most biological systems operate as a dynamic model; “cancer” occurs when the development of cancer cells exceeds the immune system’s ability to kill them. Dynamic models generally describe rates of change using derivatives. “Causes” are often modelled as affecting the derivative, and causal loops are used to indicate that variables can cause each other over time. One partial harmonizing approach “unfolds” the causal loop and defines variables according to the time they are measured. We expand on this approach. First, we explicitly distinguish between the data generating process that describes the world as it is, and the data generating process as it would be in the presence of a new intervention. Second, we leverage previous work using causal DAGs with nodes representing derivatives to allow estimation for a new range of questions previously only addressed by dynamic models. Third, our modifications are more flexible than current dynamic models because they allow for the derivatives to change over time, which may occur when interventions alter the equilibrium state.