Do Instrumental Variables Belong in Propensity Scores?

Abstract

Propensity score matching is a popular way to make causal inferences about a binary treatment in observational data. The validity of these methods depends on which variables are used to predict the propensity score. We ask: "Absent strong ignorability, what would be the effect of including an instrumental variable in the predictor set of a propensity score matching estimator?" In the case of linear adjustment, using an instrumental variable as a predictor variable for the propensity score yields greater inconsistency than the naive estimator. This additional inconsistency is increasing in the predictive power of the instrument. In the case of stratification, with a strong instrument, propensity score matching yields greater inconsistency than the naive estimator. Since the propensity score matching estimator with the instrument in the predictor set is both more biased and more variable than the naive estimator, it is conceivable that the confidence intervals for the matching estimator would have greater coverage rates. In a Monte Carlo simulation, we show that this need not be the case. Our results are further illustrated with two empirical examples: one, the Tennessee STAR experiment, with a strong instrument and the other, the Connors' (1996) Swan-Ganz catheterization dataset, with a weak instrument.