Estimating the mean and standard deviation of environmental data with below detection limit observations: Considering highly skewed data and model misspecification.
Chemosphere
; 138: 599-608, 2015 Nov.
Article
en En
| MEDLINE
| ID: mdl-26210025
ABSTRACT
In environmental studies, concentration measurements frequently fall below detection limits of measuring instruments, resulting in left-censored data. Some studies employ parametric methods such as the maximum likelihood estimator (MLE), robust regression on order statistic (rROS), and gamma regression on order statistic (GROS), while others suggest a non-parametric approach, the Kaplan-Meier method (KM). Using examples of real data from a soil characterization study in Montreal, we highlight the need for additional investigations that aim at unifying the existing literature. A number of studies have examined this issue; however, those considering data skewness and model misspecification are rare. These aspects are investigated in this paper through simulations. Among other findings, results show that for low skewed data, the performance of different statistical methods is comparable, regardless of the censoring percentage and sample size. For highly skewed data, the performance of the MLE method under lognormal and Weibull distributions is questionable; particularly, when the sample size is small or censoring percentage is high. In such conditions, MLE under gamma distribution, rROS, GROS, and KM are less sensitive to skewness. Related to model misspecification, MLE based on lognormal and Weibull distributions provides poor estimates when the true distribution of data is misspecified. However, the methods of rROS, GROS, and MLE under gamma distribution are generally robust to model misspecifications regardless of skewness, sample size, and censoring percentage. Since the characteristics of environmental data (e.g., type of distribution and skewness) are unknown a priori, we suggest using MLE based on gamma distribution, rROS and GROS.
Palabras clave
Texto completo:
1
Colección:
01-internacional
Banco de datos:
MEDLINE
Asunto principal:
Modelos Estadísticos
/
Ambiente
/
Límite de Detección
Tipo de estudio:
Diagnostic_studies
/
Prognostic_studies
/
Risk_factors_studies
Idioma:
En
Revista:
Chemosphere
Año:
2015
Tipo del documento:
Article