SIMEX for Weighting and Matching Applications with Error-Prone Covariates

• Activity Number: 3
• Type: Invited
• Date/Time: Sunday, August 3, 2014 : 2:00 PM to 3:50 PM
• Sponsor: Social Statistics Section
• Author(s): Daniel McCaffrey*+ and J.R. Lockwood
• Companies: Educational Testing Service and Educational Testing Service
• Keywords: causal effects ; nonresponse bias ; measurement error ; propensity scores

Abstract

Weighting is widely used to estimate population means from nonrandom samples. Examples include controlling for noresponse and estimating causal effects from observational studies. However, standard methods can be biased if covariates required for missing data to be ignorable are measured with error. We apply the Simulation-Extrapolation (SIMEX) method to estimate population means via weights estimated with error prone variables. We simulated the data sets according to the SIMEX procedures. On each data set, we applied the naïve approach of estimating the weighted mean of the respondents using weights equal the reciprocal of the propensity score estimated with the error prone covariates. We then fit a quadratic or quartic extrapolation curve for the projecting to zero measurement error. For causal effects, we projected the treatment effects estimated using weighting or matching. In a large simulation study, we find that compared with a naïve approach, SIMEX greatly reduces the bias in the estimated mean, but inflates the standard errors with small samples especially with the quartic extrapolation. For samples of 1000 or more, SIMEX with quadratic extrapolation is more accurate.