Primary Submission Category: Sensitivity Analysis
Robustness of Proximal Inference
Authors: Cory McCartan, Melody Huang,
Presenting Author: TBD TBD*
Proximal inference has been proposed as an alternative identification approach to relaxing traditional selection-on-observables (SOO) assumptions (i.e., no unobserved confounding) in observational causal inference. Instead of assuming researchers measure all relevant confounders, proximal inference assumes researchers have access to two informative proxies: a treatment proxy and an outcome proxy, which satisfy certain conditional independence assumptions. We formalize the trade-offs made between using a traditional SOO identification strategy in contrast to the proximal assumptions and derive the necessary scope conditions for proximal inference to provide more robust estimates than SOO. We consider the realistic setting in which both SOO and proximal assumptions are violated, finding that under even small violations of selection-on-observables, small violations in the exclusion restriction can amplify the resulting bias from proximal inference. We extend classical results from the instrumental variables literature to the proximal inference setting, and find that weak proxies can exacerbate both efficiency loss and potential bias. We compare the different approaches on a re-analysis of the impact of vote share shifts on legislative behavior.