|
Abstract:
|
Propensity score methods are popular in biomedical and policy studies, where covariates are often measured with error. However, this violates the strong ignorability assumption and threats the validity of average causal effect (ACE) estimation. Ignoring such error may lead to biased ACEs. Battistin and Chesher's bias correction formula provides reasonable ACE estimates for large samples with relatively small errors. Under an additive and nondifferential measurement error structure with relaxed strong ignorability assumption, MaCaffrey's inverse probability weighting approach provides consistent ACE estimates, but it is computationally intensive. Webb-Vargas and Stuart's multiple imputation-based approach uses external calibration data for measurement error correction, then plug in for ACE estimates. Under an additive and nondifferential measurement error structure, our group proposed a flexible Bayesian ACE estimation method based on finite mixture model without external calibration data. It provides valid ACE inference using stochastic subclassification of the underlying true latent propensity score that captures the uncertainty from measurement error to ACE estimation.
|
