A1317
Title: Handling endogenous regressors in quantile regression models: Copula approach without instruments
Authors: Rouven Haschka - Zeppelin University Friedrichshafen (Germany) [presenting]
Abstract: Endogeneity in quantile regression models has not yet received much attention in the literature. In order to handle regressor endogeneity, only instrument-based approaches are used. Seeing that instruments are often weak or unavailable, this article proposes a generalisation for the instrument-free Gaussian copula-based endogeneity correction to quantile regression models. For this purpose, two estimators are developed. First, a full maximum likelihood estimator based on directly maximising the likelihood derived from the joint distribution of explanatory variables and errors given the assumption that errors follow an asymmetric Laplace distribution. Second, a Bayesian estimator that is based on a decomposition of errors into an unobserved exponential variable and a structural normal part. By assuming that endogeneity comes from the normally distributed part, the copula endogeneity correction by control functions is used so that the model can be estimated by efficient Gibbs sampling. Identification assumptions are derived and shown under which conditions these are fulfilled. Moreover, Monte Carlo simulation results are provided to examine and compare the finite sample performances of the two abovementioned estimators and demonstrate their superiority to instrumental variable estimation in quantile regression models.
