The Impact of Measurement Error on Evaluation Methods Based on Strong Ignorability

When selection bias can purely be attributed to observables, several estimators have been discussed in the literature to estimate the average effect of a binary treatment or policy on a scalar outcome. Identification typically exploits the unconfoundedness of the treatment, which is verified if the participation status is independent of potential outcomes conditional on observable covariates. Assuming unconfoundedness, the average effect of the treatment can be estimated by matching, differencing within subpopulation averages of treated and untreated units, or by propensity score methods under an additional condition on the support of the covariates exploited. The latter condition, together with unconfoundedness, makes participation into the treatment group strongly ignorable, as defined by Rosenbaum and Rubin (1983). This paper derives conditions for identification and estimation of treatment effects when observable covariates relevant to unconfoundedness are measured with error. An expression for the measurement error bias is derived, and conditions are discussed for this to be zero. A bias correction procedure is also presented, which uses non-parametric estimates of functionals of the distribution of observed covariates.