Comparison of methods to assess the accuracy of the incorporation of censored chemical data in descriptive statistical analysis of contaminated groundwate

Chemical analyses of groundwater often present data sets with censored values, i.e., below the detection limit (LOD). When the proportion of censored values is significant, descriptive (mean, median and standard deviation) or exploratory geochemical analysis may be impaired. Ignoring such data or replacing them with some predetermined value is not always the recommended alternative. Thus, the objective of this research is to investigate the applicability of four methods in estimating censored chemical data from an area with contaminated groundwater. Three statistical methods were used: parametric (Maximum Likelihood Estimation, MLE), non-parametric (Kaplan-Meier, KM) and robust (Order Regression Methods, ROS), in addition to the traditional method of direct replacement of censored data, using LOD/2. The MLE, assuming a Gaussian distribution of the data (MLE-no), yielded allowable substitution factors, close to 0.5, similarly to the traditional substitution method (LOD/2). Validation with complete datasets with the same estimation methods and considering three artificial LOD attested to the good results of MLE-no and ROS with 25% and 50% of censored data, respectively, as well as LOD/2. The first two methods are preferable to LOD/2 as they are statistically based. It is recommended in future studies that such estimation methods be combined with other geostatistical treatments to improve the spatial analysis of hydrochemical datasets.