Quantile regression in big data: A divide and conquer based strategy

Abstract

Quantile regression, which analyzes the conditional distribution of outcomes given a set of covariates, has been widely used in many fields. However, the volume and velocity of big data make the estimation of quantile regression model extremely difficult due to the intensive computation and the limited storage. Based on divide and conquer strategy, a simple and efficient method is proposed to address this problem. The proposed approach only keeps summary statistics of each data block and then can use them to reconstruct the estimator of the entire data with asymptotically negligible approximation error. This property makes the proposed method particularly appealing when data blocks are retained in multiple servers or come in the form of data stream. Furthermore, the proposed estimator is shown to be consistent and asymptotically as efficient as the estimating equation estimator calculated using the entire data together when certain conditions hold. The merits of the proposed method are illustrated using both simulation studies and real data analysis.