Abstract: We derive nonparametric sharp bounds on the population average treatment effect (ATE) and the average treatment effect on the treated (ATT) employing an instrumental variable (IV) that does not satisfy the exclusion restriction assumption (i.e., an invalid IV). This critical assumption of IV methods is usually difficult to justify in practice since it requires that the IV affects the outcome only through its effect on the treatment. We allow the IV to affect the outcome through channels other than the treatment, and employ assumptions requiring weak monotonicity of average potential outcomes within or across subpopulations defined by the values of the potential treatment status under each value of the instrument (principal strata). There are two key features of the approach we use to derive bounds on the ATE and ATT. First, we write the parameters as weighted averages of the local average treatment effects of the principal strata, and construct bounds by first bounding each of these local treatment effects. Second, we employ a causal mediation analysis framework to disentangle the part of the effect of the instrument on the outcome that works through the treatment from the part that works through other channels. This enables us to use the (invalid) instrument to bound the causal effect of the treatment on the outcome. The bounds are employed to re-analyze the effect of Medicaid insurance on health and preventive care utilization, self-reported health status, and financial strain within the Oregon Health Insurance Experiment, allowing for the possibility that the Medicaid lottery is an invalid instrument. While our qualitative results are similar to those in the literature, our analysis suggests that some of the previous estimates may be upward biased.
Friday, March 295th. 2019
3:30pm – 5:00pm
Small reception to follow in Room 426. All are invited to attend.