THE EFFECT OF EXPANDED STANDARD DEVIATION ON WESTGARD

Download Report

Transcript THE EFFECT OF EXPANDED STANDARD DEVIATION ON WESTGARD

THE EFFECT OF SETTING QC
LIMITS GREATER THAN
ACTUAL SD ON WESTGARD
MULTI-RULES
Graham Jones
Department of Chemical Pathology,
St Vincent’s Hospital, Sydney
Background
• Westgard rules are commonly used to detect changes
in assay performances.
• The Power of Error Detection of rules can be
determined from available sources1.
• These sources assume that the SD set in the QC
protocol is the actual SD of the measurement system.
• Sometimes the SD in QC charts is set wider than the
actual instrument SD (figure 1).
• Here I investigate the effect of setting QC limits
wider than the actual SDs on the power of error
detection.
1 eg www.westgard.com
QC Limits
3
2
1
0
-1
-2
QC SD = Actual SD
•Spread of QC results across
Range.
•5% of results outside +/- 2SD
-3
3
2
1
0
-1
QC SD = 3 x Actual SD
•QC results clustered near
mean.
•No results outside +/- 2SD
-2
-3
Figure 1
Hypothesis
• Setting the SD limits in the Levy-Jennings QC
chart different to the actual instrument SD limits
will change the performance of the QC protocol.
Aim
• To investigate the utility of published Power
Function Charts to determine the Power of Error
Detection when the QCSD is larger than the Actual
SD of the method.
Setting QC Limits - illustration
A
QC SD = Actual SD
•Spread of QC results across range.
•5% of results outside +/- 2SD
+2
0
-2
+2
0
-2
B
QC SD = 3 x Actual SD
•Same ASD as graph A.
•QC results clustered near mean.
•No results outside +/- 2SD
Figure 1
Terminology
• ASD - The actual SD of the measurement system
• QCSD - The “SD” Limits set in the QC package or
drawn on Levy-Jennings chart.
• Power of error detection: The likelihood that a QC
rule will trigger for any given assay drift.
• 13s - 1 result outside 3 SD
• 12s - 1 result outside 2 SD
• 22s - 2 results outside 2 SD
• 41s - 4 consecutive results outside 1 SD
• 10x - 10 consecutive results on the same side of the
mean
Methodology
Power Function Charts were developed in Microsoft Excel.
– Random variation was simulated using the random number
generator with a normal distribution (n=1000).
– Bias was simulated by adding fixed amounts to the
randomly generated numbers.
– All rules were modelled on n=2 (2 levels of QC).
– Only change in bias (not change in precision) was
modelled (thus the R4s rule was not tested).
– the 41s and 10x rules were run on data from both materials.
– The 22s rule was run within the pair of simultaneous QCs
1
0.9
0.8
0.7
10x
4.1s
2.2s
1.3s
0.6
0.5
0.4
0.3
0.2
QCSD = ASD
Error Detection
0.1
0
1 0
0.9
1
2
3
4
5
6
0.8
0.7
10x
4.1s
2.2s
1.3s
0.6
0.5
0.4
0.3
Power Function
Charts
Effect of increasing
QCSD relative to ASD on
power of error detection
of individual rules.
QCSD = 1.5 x ASD
0.2
0.1
0
10
1
2
3
4
5
6
0.9
0.8
0.7
10x
4.1s
2.2s
1.3s
0.6
0.5
0.4
0.3
90% error detection for 10x
90% error detection for 41s
90% error detection for 22s
QCSD = 2 x ASD
0.2
0.1
0
0
1
2
3
4
5
Shift (multiples of SD)
6
Figure 2
Shift at 90% ED (xSD)
Effect of Increasing QCSD relative
to ASD
9
8
7
6
10x
4.1s
2.2s
1.3s
5
4
3
2
1
0
1
1.5
2
2.5
QCSD / ASD
Notes:
Lines are not straight due to non-smoothed data in Power Function Charts.
This graph shows the changes in error detection compared to the ASD.
Figure 3
Shift at 90% ED (xSD)
Effect of Decreasing ASD relative to
QCSD
4
3.5
3
10x
4.1s
2.2s
1.3s
2.5
2
1.5
1
0.5
0
1
1.5
2
2.5
QCSD / ASD
Notes:
Lines are not straight due to non-smoothed data in Power Function Charts.
This graph shows the changes in error detection compared to the QCSD.
Figure 4
Results and Discussion 1
• Increasing the QCSD relative to the ASD changes
the Power Function Charts for individual rules
(figure 2).
• Increasing the QCSD has different effects on the
different QC rules (figures 2 and 3).
– The error detection line for 10x does not change with
different settings of QCSD.
– The error detection lines for 41s, 22s and 13s all move
to the right, indicating reduced error detection.
• Thus the main effect is on the rules which offer
earliest error detection (the within-run rules).
Results and Discussion 2
• The effects on error detection have been expressed
relative to the actual instrument SD thus far.
• The effect of improving ASD relative to a set
QCSD is shown in figure 4.
• It can be seen that the power of error detection of
all rules change (improve) as the ASD improves.
• This demonstrates that published power function
charts cannot be used to predict error detection
when QCSD is not equal to ASD, even by using the
QCSD for the prediction.
Conclusions
• Setting the QCSD different to the ASD in QC
protocols affects the performance of Westgard
rules.
• If QCSD does not equal to ASD the published
Power Function Charts do not accurately represent
the power of error detection.
• Laboratories wishing to determine the power of
error detection using published data must set
QCSD = ASD, or develop alternate methods to
determine their own error detection power.