The first part of this talk will summarize recent work on valid post-selection inference for a multi-step pipeline: we want to use data to learn the structure of a graphical model, use that graph to identify some target causal effect, and then estimate the effect along with valid confidence intervals. We achieve valid inference using a “resampling & screening” procedure. Then we will discuss ongoing work relevant to extending this pipeline when the target effect is not point-identified because unmeasured confounding cannot be ruled out. In that case, we estimate bounds using a sensitivity model and use the independence structure encoded by the graph to gain efficiency.
