Research Papers

Partial Identification of the Marginal Treatment Effect with an Invalid Discrete Instrument (Job Market Paper)

This paper proposes a method to partially identify the marginal treatment effect when the instrument is both discrete and invalid. In the setting of a binary instrument, I employ the method of causal mediation analysis to derive nonparametric sharp bounds on local average treatment effects for different subgroups of the sample under two monotonicity assumptions. Then, applying the property that local average treatment effects can be expressed as weighted integrals of the marginal treatment effect, I obtain the sharp identification set of the marginal treatment effect under a polynomial functional form assumption. I also conduct an extension tightening the bounds on local average treatment effects by conditioning on and weighting by covariates, which in turn results in a more informative identification set of the marginal treatment effect. The properties of the method are assessed through a simulation study. Further, I apply my method to data from the Oregon Health Insurance Experiment without imposing that the lottery is a valid instrument. I find that the treatment effect of Medicaid coverage on emergency room visits decreases from always takers to compliers to never takers, regardless of whether one imposes the validity of the instrument. Moreover, my partial identification results suggest that at least 68% of individuals in the sample would increase their emergency room visits if covered by Medicaid, which is higher than the point estimation result assuming the instrument validity in Kowalski (2016, 2021), which is 48%.

Heterogeneous Treatment Effects and Better Treatment Assignment Rules: The Case of Korean Job Training Programs (with Seokchae Hwang, Chung Choe and Alfonso Flores-Lagunes )

Despite growing interest in treatment effect heterogeneity, few studies have systematically analyzed the heterogeneity of the effects of job training programs. This paper analyzes heterogeneous treatment effects on the probability of re-employment using rich administrative data from South Korea. We use a causal forest estimator to identify effect heterogeneity and analyze the distribution of the estimated treatment effects under an unconfoundedness assumption. The results suggest that 31% of trainees experience negative employment effects, despite the estimated average treatment effect on the treated being positive (0.029) and statistically significant. We examine the characteristics of the subgroups defined by the sign of the treatment effect of the training programs, and analyze conditional average treatment effects defined based on relevant covariates. We estimate practical training assignment rules informed by the estimated heterogeneous effects, as well as optimal rules using a formal machine learning algorithm. We find that alternative assignment rules based on the displaced worker's tenure and the size of the firm of previous employment perform very well. We also show that the observed assignment rule seems to suffer from substantial adverse selection into training.