A0709
Title: Parametric bootstrap inference for quantile regression and binary quantile regression
Authors: Blaise Melly - University of Bern (Switzerland) [presenting]
Martina Pons - University of Bern (Switzerland)
Abstract: The purpose is to develop (semi)parametric bootstrap inference methods for quantile and binary quantile regression models. The complete quantile regression process characterizes the entire conditional distribution of the outcome given the covariates, allowing outcomes to be resampled directly from the estimated distribution. Moreover, imposing a null hypothesis, such as a location shift, during the simulation of bootstrap samples is straightforward. This approach yields improved size and power compared to nonparametric bootstrap methods that rely on ex post recentering. The benefits of this parametric bootstrap are even greater in the context of binary quantile regression (generalized maximum score), where inference is particularly challenging due to the estimator's nonstandard asymptotic distribution and the failure of standard bootstrap techniques. By simulating outcomes from the estimated binary quantile regression process, the method provides a practical and effective approach to valid inference without requiring the selection of smoothing parameters.
