Pseudo-likelihood inference for regression models with misclassified and mismeasured variables

This paper investigates the use of a pseudo-likelihood approach for inference in regression models with covariates affected by measurement errors. The maximum pseudo-likelihood estimator is obtained through a Monte Carlo expectation-maximization type algorithm and its asymptotic properties are described. The finite sample performance of the pseudo-likelihood approach is investigated through simulation studies, and compared to a full likelihood approach and to regression calibration under different measurement error structures, as well as continuous or discrete covariates. In contrast to the full likelihood approach, our method is computationally fast while remaining competitive from an inferential perspective. Satisfactory results are also provided over regression calibration. Pseudo-likelihood and the competing methods are finally applied to the analysis of two data sets.