Multivariate analysis of a round-robin study on the measurement of chlorobiphenyls in fish oil
Abstract
Participants identified and measured chlorobiphenyls (CBs) in fish oil both before and after spiking with undisclosed amounts of four CBs following two clean-up methods, a common one and their own. Complete data were received from 18 of 30 participants of which only 2 correctly identified IUPAC No. 86. All correctly identified the other three CBs (IUPAC Nos. 52, 101 and 153). The presence of correlations (10 of 12 coefficients of correlation were moderate to high) among concentrations of the three CBs precluded univariate analysis. Following deletion of one obvious outlier, paired difference and mean data from 17 participants were analysed by a multivariate procedure. Four data pairs were compared. No effect of clean-up on precision was found. Spike recoveries differed significantly from expected values, and the patterns of differences varied with the clean-up method employed. Multivariate analysis of paired differences and means was used to rank laboratory performance with respect to laboratory precision and bias, and to identify possible outliers. Laboratory bias (between laboratory) exceeded random (within laboratory) variations in all data sets. None of the rankings based on paired differences (random error) was correlated significantly with the equivalent rankings based on paired means (laboratory bias). However, rankings for data set 1 were correlated with rankings for data set 2, as were those of data set 3 with those of data set 4 for paired difference data and for paired mean data. The multivariate method allows comparisons and laboratory ranking based on simultaneous analysis of correlated multiple determinands.