Nonparametric Dyadic Quantile Regression

Abstract

This paper considers nonparametric quantile regression for directed dyadic data observed on the ordered pairs of a complete network with n vertices. The observations exhibit dyadic dependence: dyads with disjoint indices are independent, whereas dyads sharing one or two indices are correlated. The local linear smoothing technique is adopted to estimate the quantile regression function. The asymptotic theory reveals that convergence rates of the proposed dyadic quantile regression estimation crucially depend on whether degeneracy occurs for the underlying network generating process. A rate-adaptive standardized version of the limit distribution is derived under some mild conditions. In addition, an easy-to-implement method is proposed to estimate the asymptotic variance of nonparametric dyadic quantile regression estimation. The finite-sample performance and applicability of the proposed methodology are illustrated with Monte-Carlo simulation and a real data application to bilateral portfolio holdings.