Assessment of left-censored data treatment methods using stochastic simulation

Abstract

The paper evaluates the influence of size series, percentage of censored data, and coefficients of variation used to generate synthetic series on the estimation of means, standard deviations, coefficients of variation, and medians in series with censored data. Seven techniques were applied to treat censored data in synthetic series with 180 scenarios (four size series, nine censoring percentages and five coefficients of variation): values proportional to the DL: zero, DL/2, DL/20.5 and DL - and parametric (MLE), robust (ROS) and Kaplan-Meier methods. Predictions were analyzed with four performance metrics (MPE, MAPE, KGE, and RMSE). It is found that the percentage of censored data and the coefficient of variation significantly alter forecast quality. It is also found that substitution by DL/2, by DL/20.5 and ROS are the most appropriate techniques for estimating the variables described, emphasizing ROS when estimating parametric variables and substitution by DL/20.5 for medians.