Hyewon Kim, Cornell University  PhD Candidate- Practice Job Market Talk

Title: Estimating Treatment Complementarity

Abstract: How can we estimate the complementarity between two treatments when assignment is not fully random, such as in randomized experiments with imperfect compliance or in quasi-experimental settings? The first part of this paper shows that the commonly used two-stage least squares (2SLS)—with instruments for each treatment and their interaction—is often not suitable for estimating treatment interaction effects. Specifically, 2SLS requires strong assumptions about (1) treatment effect heterogeneity and (2) types of compliers. I show that these assumptions have testable implications on first stage patterns, and these often fail in published empirical studies on complementarity. The second part of the paper proposes an alternative estimation strategy for cases where these assumptions for 2SLS are unlikely to hold. Building on the marginal treatment effect literature, this approach models potential outcomes as a linear function of individuals’ unobserved resistance to treatment and offers a clearer connection to the intended estimand of treatment interaction. Lastly, the paper revisits Angelucci and Bennett (2024), an experimental study of complementarity under imperfect compliance, to illustrate how the proposed diagnostics and alternative estimator can enhance empirical analysis of interactions between two treatments.