Assay validation for left-censored data

Abstract

In laboratory validation studies, it is often important to assess agreement between two assays, based on different techniques. Oftentimes, both assays have lower limits of detection and thus measurements are left censored. For example, in studies of Human Immunodeficiency Virus (HIV), the branched DNA (bDNA) assay was developed to quantify HIV-1 RNA concentrations in plasma. Validation of newer assays, such as the RT-PCR (reverse transcriptase polymerase chain reaction) involves assessing agreement of measurements obtained using the two techniques. Both bDNA and RT-PCR assays have lower limits of detection and thus new statistical methods are needed for assessing agreement between two left-censored variables. In this paper, we present maximum likelihood and generalized estimating equations approaches to evaluate agreement between two assays that are subject to lower limits of detection. The concordance correlation coefficient is used as an agreement index. The methodology is illustrated using HIV RNA assay data collected as part of a long-term HIV cohort study.

A nonparametric likelihood test for detecting discordance between two measurements with application to censored viral load determinations

“…Since H 0 implies equal marginal means, and the paired t - test and the Wilcoxon signed rank test are conventional for testing equal means of paired measurements, we compare the NPL test with them. For the t and Wilcoxon tests, values of X and Y below C are set equal to C , a naive (ad hoc) approach to deal with left censored data [15]. The empirical powers and significance levels are in Table 2.…”

“…First, it is hard to decide a threshold in the null hypothesis, such that the agreement coefficient above the threshold corresponds to substitutability due to the above mentioned drawbacks. Second, if the data are censored, agreement coefficient estimates can be biased, unless correctly specified parametric distributions are assumed[15]. In practice, it is not possible to know if the assumed parametric distribution applies below the detection limit.…”

A Generalized Concordance Correlation Coefficient Based on the Variance Components Generalized Linear Mixed Models for Overdispersed Count Data

“…Some robust approximations have been proposed (King and Chinchilli, 2001a), and the CCC has been recently defined and/or modified for some of these nonstandard conditions. Some examples of these are in the case of measuring time to event (Liu et al, 2005), left‐censored data (Barnhart et al, 2005), univariate censoring (Guo and Manatunga, 2007), and longitudinal repeated measurements (King, Chinchilli, and Carrasco, 2007; Carrasco, King, and Chinchilli, 2009). King and Chinchilli (2001b) generalized the expression of the CCC to allow either quantitative or qualitative data in a distribution‐free setting.…”

“…3.2.1 The total GCCC Based on the model defined in subsection 3.1, several types of CCCs can be characterized: total, intra, and inter CCC (Barnhart, Song, and Lyles, 2005; Lin, Hedayat, and Wu, 2007). The common CCC that is also known as a measure of total agreement can be expressed as an intraclass correlation between any reading from different observers on the same subject (Carrasco and Jover, 2003; Lin et al, 2007), thus where and stand for the marginal (over subjects and observers) covariance of the readings taken on the same subject and the marginal variance of data.…”

Assessing Assay Variability of Pesticide Metabolites in the Presence of Heavy Left-Censoring

“…The individual WCVs for these pairs will be zero, which in turn will lead to biased estimation for the overall WCV. Therefore, maximum likelihood estimation (MLE) which has been increasingly used to account for the left-censoring present in concentration data (Jacqmin-Gadda et al 2000; Lyles et al 2001a,b; Thiébaut and Jacqmin-Gadda 2004; Barnhart et al 2005) is a compelling alternative for estimating assay variability.…”