With only observational data on two variables, and without other assumptions,
it is not possible to infer which one causes the other. Much of the causal
literature has focused on guaranteeing identifiability of causal direction in
statistical models for datasets where strong assumption