Limit of detection values in data analysis: Do they matter?

Abstract

Abstract. Data sets containing values below the limit of detection (LOD) are known as 'censored data sets'. Such data sets are encountered regularly in most fields of environmental contaminant research. The current norm within environmental radioactivity research is to use substitution methods when analysing data sets that include values below the LOD, commonly replacing each LOD value with a value equal to half the LOD (LOD/2). However, this approach has no statistical basis and has implications when summarising or comparing data sets because it can lead to underestimates or overestimates of both the mean and the standard deviation. To remove the need to apply substitution methods, over the last four decades other fields of environmental science have been adopting statistical techniques developed for medical research applications. Despite the long history of applying these techniques in other fields and two recent environmental radioactivity publications that have used survival analysis techniques, there still seems to be reluctance within the environmental radioactivity research community to adopt these 'new' methods. This paper introduces the statistical techniques that can be used in place of LOD substitution, presents some guidance on the applicability of these techniques for different levels of data censoring and provides some examples of the use of these methods in various contexts. It is hoped the present paper will contribute to the evidence-base supporting the use of survival analysis within the field of environmental radioactivity research and go some way to changing the current norm of substitution using LOD/2.

Trace Element Distributions in the Zn-Pb (Mississippi Valley-Type) and Cu-Ag (Kupferschiefer) Sediment-Hosted Deposits in Poland

“…Datasets containing values below the limit of detection (LD) can lead to underestimates or overestimates of both the mean value and standard deviation, and therefore to correct this influence the statistical substitution methods was adopted [69] which is commonly used to replace a value below minimum limit of detection with a value equal to half the limit of detection (LD/2). In our bulk-rock geochemistry statistics we consider sample population of specific element only in the case when the number of samples with values above low detection limit constitute at least 75% and other values were treated as LD/2.…”

Early increase in anti-inflammatory biomarkers is associated with the development of multiple organ dysfunction syndrome in severely injured trauma patients

“…Statistical analysis was done using SPSS Statistics V.25 (IBM), graphs were made employing PRISM V.8 (GraphPad, California, USA). Values of biomarkers below the detection value were rated as half of the lowest detectable value 32…”

Estimation in generalized linear models under censored covariates with an application to MIREC data

“…with respect to * using an iterative equation similar to (13). Note that, as we consider only two covariates in the binary model (10), it is not very tedious to find the exact maximum likelihood estimates by evaluating the integrals within the iterative equations using a numerical integration technique.…”

“…Helsel 3 reviewed a number of existing methods for dealing with censored observations commonly encountered in clinical and environmental studies, which include the nonparametric Kaplan-Meier method for determining descriptive statistics and regression on order statistics for imputing the nondetects. Analysis of data subject to detection limits has been extensively studied in the literature in recent years (eg, other works [4][5][6][7][8][9][10][11][12][13] ). Helsel 4 reviewed existing statistical methods for dealing with nondetects in environmental data.…”