Learning and using graphical structure for reliable estimation of causal effects from observational studies

It is commonplace in observational health studies to use causal graphical models, usually directed acyclic graphs (DAGs), to inform the estimation of causal effects. Most often, a single DAG or small set of DAGs is posited based on domain knowledge and these graphs are used to inform the selection of adjustment variables, i.e., possible confounders for the effect of interest. What if the posited graph is wrong? Causal discovery algorithms use the data itself to select among possible DAGs or causal structures. The “discovered” graph may then be combined with causal effects estimators to produce effect estimates that are “data-driven” in the sense that they are less reliant on controversial or possibly misspecified model assumptions. Several questions and statistical challenges arise in combining causal discovery algorithms with the estimation of causal effects: this talk will propose an approach to valid post-selection inference and exploiting structure to more efficiently estimate bounds when unmeasured confounding cannot be ruled out.