We consider both a parametric and a semiparametric method to account for classification errors on the dependent variable in an ordered response model. The methods are applied to the analysis of self-reported speaking fluency of male immigrants in Germany. We find that a parametric model which explicitly allows for misclassification performs better than a standard ordered probit model and than a model with random thresholds. We find some substantial differences in parameter estimates and predictions of the different models.
展开▼
