Primary Submission Category: Instrumental Variables
Synthetic Instrumental Variables in Diff-in-Diff designs with unmeasured confounding
Authors: Jaume Vives-i-Bastida, Ahmet Gulek,
Presenting Author: Jaume Vives-i-Bastida*
Unmeasured confounding and selection into treatment are key threats to reliable causal inference in Difference-in-Differences (DiD) designs. In practice, researchers often use instrumental variables to address endogeneity concerns, for example through shift-share instruments. However, such instruments may be correlated with unobserved confounders, exhibiting pre-trends. This paper explores the use of synthetic controls to address unmeasured confounding in IV-DiD settings. We propose a synthetic IV estimator that partials out the unmeasured confounding and derive conditions under which it is consistent when the standard two-stage least squares is not. Motivated by the finite sample properties of our estimator we then propose ensemble estimators that might address different sources of bias simultaneously. Finally, we show the relevance and pitfalls of our estimator in a simulation exercise and in an empirical application.