Propensity Score Methods When Covariates Are Measured with Error

Propensity score methods are commonly used to estimate causal effects in non-experimental studies in prevention science. These approaches are particularly helpful when the researcher cannot randomize on the variables of interest. However, there are some design challenges associated with the use of propensity scores. For example, existing propensity score methods assume that covariates are measured without error but covariate measurement error is likely common, especially in many prevention science research studies where constructs such as depression, childhood maltreatment, or even socio-economic status cannot be measured directly. The purpose of this talk is to suggest an alternative propensity score matching approach which addresses some of the issues associated with covariate measurement error.

First, we will provide an overview of the use of propensity scores in prevention science research and highlight common errors researchers make when trying to employ propensity score methods. Then, we will focus on the implications of measurement error in the covariates on the estimation of causal effects using propensity scores, with a focus on topics relevant for prevention science. The talk will introduce a method called Multiple Imputation using External Calibration (MIEC), which can account for covariate measurement error in propensity score estimation. MIEC uses a main study sample and a calibration dataset that includes observations of the true covariate (X) as well as the version measured with error (W). MIEC creates multiple imputations of X in the main study sample, using information on the joint distribution of X, W, other covariates, and the outcome of interest, from both the calibration and the main data. We will provide a summary of some examples of the use of MIEC. We will then summarize some findings from a series of simulation studies we conducted in which we found that MIEC estimates the treatment effect almost as well as if the true covariate X were available. This series of studies also indicated that the outcome must be used in the imputation process, a finding related to the idea of congeniality in the multiple imputation literature. We will then illustrate MIEC using an example estimating the effect of neighborhood disadvantage on the mental health of adolescents, where the method accounts for measurement error in the adolescents' report of their mothers' age when they (the adolescents) were born. We will conclude by summarizing some advantage of MIEC over other traditional matching approaches. Recommendations will be made for future use of MIEC in prevention science studies.