The impact of incomplete data on quantile regression for longitudinal data

Abstract

We investigate the performance of methods for estimating the conditional quantile of a response based on longitudinal data, when outcomes are incomplete and when the correlation between repeated responses is ignored. In a simulation study, we compare the performance of the quantile regression estimator based on the complete cases, the available cases, quantile-based multiple imputation, and quantile-based inverse probability weight-ing. In the data setting considered, quantile-based multiple imputation is the most promising method with the best bias-efficiency trade-off. A potential drawback, however, is its computation time.