GUN. CHEM.
25/6, 863-869 (1979)
Power Functionsfor StatisticalContro!Rules
James 0. Westgard1 and Torgny Groth2
We have studied power functions for several
control
rules
by use of a computer simulation program. These power
functions show the relationship between the probability
for rejection and the size of the analytical errors that are
to be detected. They allow some assessment of the quality
available from present statistical control systems and
provide some guidance in the selection of control rules and
numbers of control observations when new control sys-
tems are designed.
In an earlier report (1), we assessed the relative performance
of many different statistical control rules on the basis of their
probabilities for false rejection and error detection. We
identified rules with high probabilities for false rejection and
those most sensitive to different kinds of analytical errors.
one rule provided the best performance for all criteria,
judged by a low probability for false rejection and a high
probability for detecting both random and systematic errors.
In selecting a control system, we recommended that the rules
with high probabilities for false rejection be eliminated from
consideration, and that a combination of the remaining rules
be chosen to include at least one rule sensitive to random error
and one sensitive to systematic error. An example of a control
system involving a combination of Shewhart and cumulative
sum rules was presented in detail to illustrate the potential
improvements made possible by a careful choice of control
rules (2).
Although this earlier work has been useful in understanding
the relative performance of different control systems, it does
not provide the complete information needed for a quantita-
tive assessment of the quality being achieved by present
control systems. To do this, more information is needed about
the probability for error detection over a wide range of random
and systematic errors. In the present study we have further
investigated, via computer simulation, the performance of the
most sensitive control rules. The performance of each control
rule is summarized graphically by representation of a power
function, which is a plot of the probability for rejection vs. the
size of the analytical error. These power functions should be
useful both for assessing the quality of existing control systems
and for selecting control rules and the numbers of control
observations when designing new control systems.
I Departments of Medicine, Pathology, and Clinical Laboratories,
University of Wisconsin Center for Health Sciences, Madison, WI
53706.
2The Group for Biomedical Informatics, Uppsala University Data
Center, S-750 02 Uppsala, Sweden.
Received June 1, 1978; accepted March 6, 1979.
Methods and Materials
Definitions
ofTerms
The probability for rejection, p, is the probability that a
given control rule will generate a rejection signal. When no
analytical errors are present except the inherent imprecision
of the analytical method, the probability for rejection can be
described as the probability for false rejection (pfr). When an
analytical error is present, the term probability for error de-
tection (Ped) can be used.
LRE is the size of the random error to be detected. It is
given as a factor, such that a IRE of 2.0 means that the
standard deviation, s, of the method has doubled. This term
differs from the convention used in the earlier study (1), where
No increases in random error were presented as percentage in-
as creases in the standard deviation. Thus, a 100% increase in
random error would be the same as a RE of 2.0.
LSE is the size of the systematic error to be detected. It is
convenient to express SE in multiples of the standard de-
viation of the analytical method. This avoids the use of dif-
ferent concentration units and permits presentation of the
power functions in a more generally useful form. A L.SE of 2.Os
will mean a systematic error equivalent to two times the
standard deviation of the analytical method.
“Power function” refers to a way of presenting information
about the performance of a control rule by a plot of the
probability for rejection (p) vs. the size of the analytical error
(either random, S.RE, or systematic, L.SE). To relate the
power function for a particular control rule to the character-
istics presented earlier (1), note that the probability for false
rejection is given by the y-intercept, i.e., the y-value (or
probability for rejection) when the x -value (or error level) is
zero. The probability for error detection is given by they-value
corresponding to the x-value or error level of interest. In the
earlier study, the error levels were RE = 2.0 and ISE = 1.Os.
The power functions have been prepared by performing
computer simulations for many different error levels, in the
manner described earlier (1). The cost of the simulations made
it impractical to perform this complete a study on the per-
formance of all of the control rules selected initially.
Control rules are identified by symbols of the form AL,
where A is an abbreviation of a statistic or the number of
control measurements, and L is the control limit. For example,
1 refers to the Shewhart control rule, in which an analytical
run is declared out of control when a single control measure-
ment exceeds the mean .O9 standard deviations. All the
symbols used here are defined in more detail in our earlier
paper (1 ).
CLINICALCHEMISTRY,Vol. 25, No. 6, 1979 863
 12$ CONTROL RULE 22$ CONTROL RULE

1.0- 1.0-

0.8-
a. a.

>-
I- 0.6- F- 0.6-
-4

-J -J
1

0.4- 0.4-
a a
C C
0.2- 0.2-
a-

0.0- 0.0-

I
0.0
I
1.0
I
2.0
1
3.0
i
4.0
0.0 1.0 2.0 3.0 4.0

SE (S) ASE IS)

Fig. 1. Power of the 128 control rule for detecting systematic Fig. 3. Power of the 223control rule fordetecting
systemaUc
error error
Theprobabilityfor rejection (p), y, is plotted vs. the size of the systematic error
(SE). x. For meaning of symbols in the aphed lines, in all Figures, see
text.

Is/22s CONTROL RULES

1.0-
‘3S CONTROL RULE
‘‘ 0.8-
1.0-

a.

o.e-
>-
0. 0.6-

>- -J
0.6-
0.4-
-J
-4 co
0.4- 0
a a- 0.2-

0
0.2-
a- I I I I I
0.0-

0.0 1.0 2.0 3.0 4.0

0.0-

oo 1o 2!O SE (S)
Fig. 4. Power of the 13/228 combination of control rules
for
5E (5)
detecting systematic error
Fig. 2. Power of the 138 control rule for detecting systematic
error
for false rejection (pfr). Point-to-point curves were drawn by
computer; irregularities in the lines represent the uncertainty
in the simulation results. Using these power fupctions will
Computer Simulation Program necessitate some visual averaging and interpolation.
This has been described elsewhere (1). Four hundred ana-
lytical runs were simulated for systematic errors (SE) of
Results
O.25s, O.50s, O.75s, 1.OOs, 1.50s, 2.OOs, 3.OOs, and 4.OOs, and for For the different control rules considered here, Figures 1-10
random errors (ARE) of 1.10, 1.25, 1.5O 1.75, 2.00, 2.50, and present power functions for detecting systematic error. Fig-
3.00. Sixteen hundred analytical runs were simulated for the ures 11-20 present power functions for detecting increases in
baseline conditions (SE = 0, RE = 1.0). This point falls on random error. The several curves presented in each figure
the y-axis for each curve and corresponds to the probability correspond to different numbers of control observations (N

804 CLINICAL CHEMISTRY, Vol. 25, No. 6, 1979

 13s/22S/41s/IOg CONTROL RULES 43sf s)t .7S CONTROL RULES
1.0- 1.0-

0.8- ‘‘ 0.8-
a.

>-
0.6- 0.6-
I-
-J -J

0.4- cn
a
C C
a:
a- 0.2- a. 0.2-

0.0- 0.0-

0.0 1.0 2.0 3.0 4.0 0.0 1.0 2.0 3.0 4.0

SE IS) SE IS)
Fig. 5. Power ofthe 1/2/4/1O combination of control rules
FIg. 7. Power of the 1/CS1 combination of control rules for
for detecting systematic error detecting systematic error

CSI#{176}7CONTROL RULE xo.05 CONTROL RULE

1.0- 1.0-

0.6- 0.8-
0. 0.

0.6- >-
I- 0.6-
I-
-J -J
-I

CD 0.4- ai 0.4-
a a
CD
C 0
a:
a- 0.2-

0.0- 0.0-

20 3!0 4O 1O 2!O 3!0 4O

SE IS) SE (S)
FIg. 6. Power of the CS1 control rule for detecting systematic Fig. 8. Power of the x005control rule for detecting systematic
error error

= 1,2,4,8, 12, and 20, as indicated by the symbols +, D, V, ,
0, and , respectively). The top curve in each figure will
correspond to N = 20, and the lowest curve to N 1, assuming
that the control rule can be invoked with one control obser-
vation. For control rules that require more than one obser-
vation, such as the mean or range rules, the lowest curve will
correspond to N = 2.
Discussion
In describing the performance characteristics
tests, it has been common in the literature
or type I error, and 3, or type II error. The power of a statis-
tical test is expressed as I minus the probability for a type H
error (3). This is actually the probability for rejection, as de-
termined in our studies. Power functions are piots of the power
for rejection vs. the size of the error to be detected.
There has been little discussion about the power of the
control systems in use in clinical chemistry laboratories.
Larsen et al. (4) provided figures for average run length
(ARL) for a decision limit cusum control system. ARL gives
of statistical information about the power of the control system in that
to use the terms a, 1/ARL is equal to the probability for rejection (5). Statistical

CLINICAL CHEMISTRY. Vol. 25, No. 6, 1979 865

 CONTROL RULE 12S CONTROL RULE
1.0-

0.8- 0.6-
0. 0.

>- >-
0.6- 0.6-
I- I-
-J -J

0.4- 0.4-
a a
m
0 C
0:: c
a- 0.2-
3- 0.2-

0.0- 0.0-

0.0 1.0 2.0 3.0 4.0 2!O 2S 3!0

SE (S) REvs. the increase In random error

The probability for rejection (p), y, is plotted
Fig. 9. Power of the control rule for detecting systematic Fig. 11. Power ofthe 12s rule for detecting
control random
error error
(ARE), x

R00/R00 CONTROL RULES

1.0- 13$ CONTROL RULE

1.0-

0.8-
a
0.8-
a.
0.6-
I.-
>-
-J 0.6-

0.4- -J
a
0.4-
C a
0.2-
0 C
0.2-

0.0-#{149}

0O 10 20 3!0 4O 0.0-

2O 30
SE. (Si
Fig. 10. Powerof the x001/R0,01 combination ofcontrol rules for RE
detecting systematic error
FIg. 12. Power of the 13 controlrule for detecting random

and quality-control texts sometimes give figures for ARL and
sometimes present operating characteristic curves, or OC
curves (6). OC curves provide similar information about the
power of the statistical tests, but in an inverse manner. They
present the probability of a control observation falling within
the control limits, rather than the probability for exceeding
control 1imit (which is the probability for rejection, or the
power of the test).
We think the form of the power function as presented here
describes the statistical properties of the control rules in a
manner easily used by laboratory analysts. For example, as-
error
sume that an existing control system uses the 1 control rule
with two control observations per analytical run. What sys-
tematic error can be detected with a probability of 0.50 (or 50%
chance)? To determine this, refer to Figure 2, which shows the
power functions for the 1 control rule (for systematic error).
The second curve from the bottom gives the information for
two control observations. The 0.50 probability point on this
curve (y-value or y-coordinate) corresponds to an x-value or
SE of 2.5s. This means that there is a probability of 0.50 (or

866 CLINICALCHEMISTRY, Vol. 25, No. 6, 1979

 i LOS
22S CONTROL RULE 13S/CS CONTROL RULES
1.0- 1 .0-

0.6- o.e-
Q. a.

>- >-
0.6- 0.6-
I-
-J -j
I-,

0.4- 0.4-

C C
a: 0.2- 0.2-
0 ii

0.0- 0.0-

1.0
I
1.S 2.0
I 2.S
I I
3.0 1.0 1.S 2.0 2.S 3.0

RE RE

rule for detectingrandom

Fig. 13. Power of the 22$ control Fig. 15. Power of the 13/CS1 combination of control rules
error for detecting random error

3S/22S CONTROL RULES R005 CONTROL RULE

1 .0- 1 .0-

0.8- 0.8-
a. a.

>- >-
I.-. 0.6- 0.6-
I-
-
-J
co 0.4- co 0.4-
a
-
0 0
c c 0.2-
O.2 0

0.0- o.o-

1O 2O 2 3!O 2O 21.

RE RE

Fig. 14. Power of the 1/223 combination of control rules for Fig. 16. Power of the R005 control rule for detecting random
detecting random error error

50% chance) of detecting a systematic error equivalent to 2.5
times the standard deviation of the analytical method.
As another example, assume that the objective is to select
a control system to maintain a specified level of quality. What
control system will be able to detect a systematic error
equivalent to 3.4 times the standard deviation of the analytical
method (zSE = 3.4s) with a probability of 0.90? From Figure
2, this error could be detected with the desired probability by
the 138 rule with three or four observations. From Figure 4,
however, this error could be detected by the 1a/2 combi-
nation of rules with N = 2 (two observations); thus, this
combination provides the desired probability for error de-
tection with fewer control observations. In a similar manner,
one could inspect all the power functions, then choose the
appropriate control system based on the lowest N, the lowest
probability for false rejections, the simplest data calculations,
or other practical considerations.
Power functions, therefore, should be useful in assessing
the performance of existing control systems and also in the
design of new and improved control systems. If the design of
a control system were to relate the choice of control rules and
the number of control observations to some quality specifi-

CLINICALCHEMISTRY, Vol. 25, No. 6. 1979 867

 R001 CONTROL RULE X#{149}05CONTROL RULE
1.0- 1.0-

#{149}‘0.6- 0.8-
a.

>- >-
0.6- F- 0.6-
-4
-
-J
p-I

m 0.4- 0.4-
a
-

0 0
a: 0.2-
a. 0

0.0- 0.0-

2!0 2!S 3!0 1.0 i.S 2.0 2.S 3.0

RE RE

control rule for detecting random

Fig. 17. Power of the R0#{149}01 Fig. 18. Power of the X.o5 control rule for detecting random
error error

cation, there would then be a known, specified level of quality

in routine analyses.
We are studying the design process in more detaiL3’4 It is
apparent that the process will include the following steps: (a) X CONTROL RULE
defmition of a value for the allowable analytical error and the
1.0-
medical decision concentration where this performance is
critical; (b) estimation of the method’s standard deviation and
bias at that medical decision concentration by the use of data
0.8-
aquired in method performance studies; (c) calculation of the 0.
systematic error and the random error that must be detected
by the control system if it is to maintain quality within the
defined specification; (d) choice of a value for the desired >-

I- 0.6-
probability for detecting those analytical errors of interest; ‘-I
-J
and (e) inspection of power functions for control rules to de-
termine which rules and what number of control observations 0.4-
permit detection of those analytical errors of interest with the a
desired probability.
0
Computer also appears to be a useful tool in
simulation
0.2-
developing a design procedure.5 To closely simulate the
characteristics of the analytical method, an interactive sim-
ulation program is desirable, because the power functions are
affected by the size of the between-run component of variation 0.0-
and by the choice of control limits (calculated from within-run
vs. total standard deviation) (7). This makes it impractical to 2O 2S 3O
determine all the possible power functions of interest and
present them in graphic or tabular form. An interactive pro- RE
gram can permit their generation for the conditions of in-
terest. Fig. 19. Power of the X.oi control rule for detecting random
Because of the effects of a between-run component of error
variation, the power functions presented here are idealistic

Wetgard, J. 0., and Groth, T., Design of a statistical control
procedure to achieve a specified standard of analytical performance.
Clin. Chem. 24, 1049 (1978). Abstract.
4Groth, T., and Westgard, J. 0., Design of statistical control pro-
cedures utilizing quality specifications and power functions for control
rules. VHth Tenovus Workshop, Quality Control in Clinical Endo-
crinology, Tenovu8 Institute for Cancer Research, Cardiff, U.K., May
3-4, 1979.
5Groth, T., Falk, H., and Westgard, J. 0., An interactive computer
simulation program for the design of statistical control procedures
in clinical chemistry, manuscript in preparation for Computer Pro-
grams in Biomedicine.
for most applications; that is, an assessment of the quality of
an existing control system will tend to be optimistic. The error
level that may occur is likely to be greater than that predicted,
and the number of control observations specified in designing
a control procedure will likely be too small. Despite these
limitations, these power functions provide a useful first step
in assessing the performance of control systems. They allow
some revealing insights into the relationship between the
quality achieved or desired and the control rules and numbers
of control observations that must be used.

868 CLINICAL CHEMISTRY, Vol. 25, No. 6, 1979

 CONTROL RULES During the course of this work, J.O.W. was supported by project
g001100, grant 170794 from the Graduate School Research Committee, Uni-
1.0-
versity of Wisconsin-Madison. Computer support was provided by
the Uppsala University Data Center and financed by funds from the
University to the Medical Faculty. We thank Professor Cart-Henric
de Verdier for his continuing support and encouragement.
0.0-
0. References
1. Westgard, J. 0., Groth, T., Aronsson, T., et a!., Performance
characteristics of rules for internal quality control: Probabilities for
I- 0.6- false rejection and error detection. Cin. Chem. 23, 1857-1867
(1977).
-J 2. Westgard, J. 0., Groth, T., Arousson, T., and de Verdier, C-H., A
‘-.4

0.4-
combined Shewart-cusum control chart for improved quality control
a in clinical chemistry. Clin. Chem. 23, 1881-1887 (1977).
3. Dixon, W. J., and Massey, F. J., Jr., Introduction to Statistical
C Analysis, 3rd ed., McGraw-Hill, New York, NY, 1969, p 85.
a- 0.2- 4. Larsen, L. H., Williams, R D. B., and Nicol, G. R., Automated
quality control of biomedical data. Part I: System description. CRC
Crit. Rev. Clin. Lab. Sci. 8, 241 (1977).
5. Jardine, A. K. S., MacFarlane, J. D., and Greensted, C. S., Statis-
0.0-
tical Methods for Quality Control, The Pitman Press, Bath, U.K.,
1975, p 133.
1.0 1.S 2.0 2. 3.0
6. Duncan, A. J., Quality Control and Industrial Statistics, 3rd ed.,
Richard D. Irwin, Inc., Homewood, IL, 1965, p 443.
ARE 7. Westgard, J. 0., FaIk, H., and Groth, T., Influence of a between-run
component of variation, choice of control limits, and shape of error
FIg. 20. Power ofthe x0#{149}01/R0#{149}01
combination of control rules for distribution on the performance characteristics of rules for internal
detecting randomerror quality control. Clin. Chem. 25,394-400(1979).

CLINICALCHEMISTRY,Vol. 25, No. 6, 1979 869