Estimating Local Average Treatment Effects Under Weak Monotonicity
We study the estimation of treatment effects with an instrumental variable when there are compliers and defiers. If the two groups exist at distinct covariate values, then the instrumental variable identifies a local average treatment effect, but conventional two-stage least squares does not estimate it. An oracle with knowledge of the defier covariate values could correct the two-stage least squares estimate by applying a sign correction wherever defiers are present. We propose a feasible version of this estimator where estimated propensity scores are used to identify defier covariate values. We characterise the asymptotic distribution of the feasible estimator and show conditions under which it is asymptotically equivalent to the oracle estimator. We provide a consistent estimator for the standard error of the feasible estimator, allowing inference. We evaluate the finite sample properties of the feasible estimator by simulation, and find that it outperforms the oracle estimator in terms of mean squared error. We apply the estimator to the data of Card (1993). Compared to conventional two-stage least squares estimates, the point estimates obtained by our estimator differ in sign and are insignificant. These results suggest that the instrument of Card (1993) is weak.
Testing Instrument Validity with Covariates (with Toru Kitagawa)
Instrumental variable methods are widely used in causal inference, but the validity of the identifying assumptions is often disputed. Tests of these identifying assumptions exist, but are rarely used in practice as they struggle to accommodate conditioning covariates, which researchers often use to justify instrument validity. We develop a novel test of the instrumental variable identifying assumptions for heterogeneous treatment effect models that can easily accommodate conditioning covariates without compromising power. We assume semiparametric dependence between potential outcomes and conditioning covariates. This allows us to replace testable implications for the outcome variable which hold conditional on covariates with unconditional testable implications for estimable partial residuals, effectively performing dimensionality reduction. We obtain testable equality and inequality restrictions among the subdensities of partial residuals, and discuss how they can be employed to detect invalid instruments. To improve power, we propose distillation, where a trimmed sample is used to test the inequality restrictions. We show asymptotic validity of our test, and in Monte Carlo exercises we find gains in finite sample power from testing restrictions jointly and distillation. We apply our test procedure to three instruments and reject the null for one.
Instrument Validity in Two-Stage Least Squares (with Toru Kitagawa)
Coming soon!