45
| 4 | 2011
Klinisk Biokemi i Norden
same as that of the prediction of the test result from
the regression function [8].
The differences can be displayed as a “folded expe-
rimental cumulative frequency plot” or “mountain
plot” [9].To achieve this, the results are ranked and
their percentiles plotted. At the 50-percentile (median)
the percentile is reversed, i.e. the graph mirrored, or
folded, thus producing a mountain-like diagram with
the peak corresponding to the median. In figure 3the
“
mountain” has been tilted 90 º and superimposed on
the difference plot.
The difference plot shows the distribution of the dif-
ferences; vertically the difference between the results
and horizontally one of the variables or the average
of both. The mountain plot also illustrates the distri-
bution of the differences but shows their cumulative
probability which gives a visual impression of its sym-
metry and closeness to a Gaussian distribution. The
X-axis is the percentiles and since these are mirrored
at 0.5 and the “peak” will correspond to the median
of the data set.
A result that deviates much from the rest might be
suspected to be an outlier. It is important to be restric-
tive in eliminating any data and it is useful to recal-
culate everything with and without a suspected result.
If the difference is minor then it is wise to include the
suspected result.
Clinical considerations
A statistical significance of a difference between two
results is determined by the size of the difference and
the uncertainty of the measurements (1). In a clinical
context other facts need to be considered e.g. biological
variation and the risk that an erroneous result might
harm the patient. For the laboratory it is important to
decide if the difference to a previous procedure will
prompt a change of reference intervals, decision points
or other set-points.
Various measures are used to support the evalu-
ation of the comparison. Provided there is a linear
relationship between the results – and that is feasible
to assume since the same measurand is targeted – then
the percentage of results that should be within a cer-
Figure 4.
A regression analysis of two procedures for B-Haemoglobin concentration indicated a slope of 1.11 and an intercept
of -16 units. An allowable difference of3 % shows 95,1 % of the observations within the ATE (dotted lines). The cut-off limits
define the diagnostic performance of the test procedure in relation to the comparative procedure. With the chosen set-points
a high sensitivity (in relation to the comparative procedure) is obtained for anemia, whereas a high specificity is obtained for
“
hemochromatosis (?)”. The table also gives the Youden and K-indices which describe the ROC-curve.
High
Low
True pos:
4
33
True neg:
48
48
False pos:
1
13
False neg:
2
1
Sum:
55
95
Sensitivity:
0,67
0,97
Specificity:
0,98
0,79
LR:
33,5
4,6
Youden ind:
0,65
0,76
K Index:
0,33
0,21
Efficiency:
0,95
0,85
Prevalence:
0,11
0,36
PV (+):
0,80
0,72
PV (-):
0,96
0,98
L/E
E
Allow diff, %:
3
ATE LER above LER below
95,1
2,9
2,0
Abs/Rel:
Rel
(
Fortsætter side 46)
