36 |
Klinisk Biokemi i Norden · 1 2013
case of, for example, the triplicates recommended in
the EP14-A2 protocol. As a result, this study shows
the effect of a high number of measurements on redu-
cing the magnitude of the prediction intervals used
in the statistical approach. This is nicely exemplified
by the small prediction intervals at the mean of the
concentrations covered by the native sera for S-Ca
(≤0.5%) and S-Prot (≤0.7%). Note that the prediction
intervals in Table 1 at the concentration of EQA
sample #1 are very similar to those at the mean, but
not necessarily for sample #2. This is because of the
typical profile of the 95% prediction interval. There-
fore, we use here the prediction intervals at the mean
to discuss the effect of the number of measurements.
A calculation example may help to clarify the state-
ment about reducing the prediction interval with
increasing number of measurements. In this study,
the Sy/x of the regression analysis was for calcium
typically in the 0.21% range (Sy/x in % calculated
from the absolute value in mmol/L relative to the
mean concentration). Note also that in the absence of
sample-related effects, it reflects the expected varia-
tion in the method comparison due to measurement
imprecision. Using the median CV of 1% observed
in this study, we expected for a single measurement
an Sy/x of approximately 1% (assuming no error in
x). This means that to reduce the Sy/x to 0.21%, 23
measurements are needed. In this study we achieved
this number, for example, in the S-Ca Siemens Advia
peer group comprising 11 laboratories measuring
each sample in duplicate (n = 22). According to the
typical EP14-A2 protocol with 3 measurements, the
Sy/x would have been nearly 3 times bigger. The
benefit of working with narrow prediction intervals
is an increased potential to uncover subtle matrix-
effects (= non-commutability) of EQA materials. As
mentioned before, this may be necessary for certain
applications, such as trueness assessment against
quality specifications. Indeed, if an EQA material is
applied to assess the bias of, for example, a S-Ca assay,
its non-commutability should consume only 1/3 of
the bias specification derived from biological varia-
tion, i.e., 0.8%. In consequence, prediction intervals
in the order of 0.3% are needed for commutability
assessment. From this point of view, commutability
assessment from special EQA surveys with native
samples are utmost valuable.
Nevertheless, commutability studies may still be
Vandringstur på Island. Foto: Ingunn Þorsteinsdóttir
