Influência dos valores censurados na determinação da concentração média de variáveis de qualidade da água

Abstract

One of the problems encountered in the statistical analysis of water quality monitoring data is the censored data. Censoring, occurs when the value of a measurement or observation is only partially known, it is below or above the detection limit (DL) of analytical measurement technique. It is common in the literature replacing censored values by zero or a fraction (1⁄2DL, or DL) of the detection limit for each nondetect. Two decades of research has shown that if the censored data is not handle properly lead to inaccurate estimates of statistical parameters (mean and standard deviation) in comparison with other methods, especially when the data set has small number of observations (n <50) or a high percentage of censored data (above 15% of total observations). In this context the goal of this study was to highlight how the mean concentration of water quality variables are influenced by sample size, percentage of censored data and calculation method. The study clearly illustrate the problems with substitution of arbitrary values for nondetects by testing them against parametric methods Maximum Likelihood Estimator (MLE) and Regression on Order Statistics (ROS) and nonparametric method Kaplan-Meier (KM), designed expressly for dealing with censored data. The analysis was carried out with dataset from Cuiabá River water quality monitoring program. The results show that substituting methods should be avoided because sample means and standard deviations are inaccurate and irreproducible.