Covariate measurement error: bias correction under response-based sampling

Abstract

In this paper we propose a general framework to deal with the presence of covariate measurement error (CME) in response-based (RB) samples. Using Chesher's (1991) methodology, we obtain a small error variance approximation for the contaminated sampling distributions that characterize RB samples with CME. Then, following Chesher (2000), we develop the generalized method of moments (GMM) estimators that reduce the bias of the most well known likelihood-based estimators for RB samples which ignore the existence of CME and derive a score test to detect the presence of this type of measurement error. Our approach only requires the specification of the conditional distribution of the response variable given the latent covariates and the classical additive measurement error model assumption, the availability of information on both the marginal probability of the strata in the population and the variance of the measurement error not being essential. Monte Carlo evidence is presented which suggests that, in RB samples of moderate sizes, the bias-reduced GMM estimators perform well.