Come and be part of our next seminar discussion on April 30th! Carlos García Meixide, CTML's Visiting Student Researcher will present his talk on "Causal Inference Via Proxy Interventions." The talk will take place at 12 PM at Berkeley Way West, 5th Floor, Room 5401.
Identifying causal effects by deterministically fixing an exposure leans on assumptions rarely satisfied in real-world settings. The simple case of writing a counterfactual mean in terms of the observable probability law demands that every unit—no matter its characteristics—has a nonzero chance of receiving a given treatment option. Even when this assumption holds at the population level, empirical performance degrades when it is only weakly satisfied.
This fragility in deterministic interventions opens the door to the more stable alternative of manipulating the treatment assignment process itself instead of hard fixing treatments outright. These interventions are not only mathematically better-posed, but also more realistic from a clinical viewpoint—while still recovering classical deterministic interventions as a limiting case.
We introduce the idea of intervening on lower-dimensional aspects of a nonparametric structural model using a potentially invalid instrument, rather than directly manipulating treatment assignments. These indirect stochastic interventions on the treatment through instruments yield valid causal inference under unmeasured confounding. We discuss data adaptive selection of such instruments, and the implications of these oracle dimension reductions they induce