Multiple Imputation Based on Conditional Quantile Estimation

Abstract

Multiple imputation is a simulation-based approach for the analysis of data with missing observations. It is widely utilized in many set- tings and preeminent among general approaches when the analytical method does not involve a likelihood function or this is too complex. We consider a multiple imputation method based on the estimation of conditional quantiles of missing observations given the observed data. The method does not require modeling a likelihood and has desirable features that may be useful in some practical settings. It can also be applied to impute dependent, bounded, censored and count data. In a simulation study it shows some advantage over the alternative meth- ods considered in terms of mean squared error across all scenarios except when the data arise from a normal distribution where all meth- ods considered perform equally well. We present an application to the estimation of percentiles of body mass index conditional on physical activity assessed by accelerometers.