The finite sample performance of instrumental variable-based estimators of the Local Average Treatment Effect when controlling for covariates

TL;DR

This study evaluates various instrumental variable estimators' finite sample performance for estimating local average treatment effects, highlighting the effectiveness of machine learning methods like random forests in this context.

Contribution

It provides a comprehensive simulation comparison of parametric, semi-parametric, and non-parametric estimators, emphasizing the performance of random forest-based methods in finite samples.

Findings

Random forest estimators have competitive coverage and low bias.

Kernel regression achieves the lowest root mean squared error.

Non-parametric methods like random forests are effective for covariate control.

Abstract

This paper investigates the finite sample performance of a range of parametric, semi-parametric, and non-parametric instrumental variable estimators when controlling for a fixed set of covariates to evaluate the local average treatment effect. Our simulation designs are based on empirical labor market data from the US and vary in several dimensions, including effect heterogeneity, instrument selectivity, instrument strength, outcome distribution, and sample size. Among the estimators and simulations considered, non-parametric estimation based on the random forest (a machine learner controlling for covariates in a data-driven way) performs competitive in terms of the average coverage rates of the (bootstrap-based) 95% confidence intervals, while also being relatively precise. Non-parametric kernel regression as well as certain versions of semi-parametric radius matching on the propensity score, pair matching on the covariates, and inverse probability weighting also have a decent coverage, but are less precise than the random forest-based method. In terms of the average root mean squared error of LATE estimation, kernel regression performs best, closely followed by the random forest method, which has the lowest average absolute bias.