QUALITY ENGINEERINGÕ

Vol. 16, No. 3, pp. 407–421, 2004

Robust Individuals Control Chart for Exploratory Analysis

Jeroen de Mast1,* and Kit C. B. Roes2

1
Institute for Business and Industrial Statistics of the University of Amsterdam
(IBIS UvA), Amsterdam, The Netherlands
2
NV Organon, Oss, The Netherlands

ABSTRACT

Apart from their uses in the context of statistical process control, control charts
are also used as an exploratory tool in the context of exploratory data analysis.
These two application situations are not similar and, as a consequence, control
charting methodology should be adjusted to the exploratory context. In this
article we make an inventory of the requirements for control charts that are used
for exploratory analysis and we propose a procedure that meets these
requirements. Robustness against assignable causes of variation appears to be
important. The proposed methodology is illustrated from two real-life examples.

Key Words: A-estimator; Change-point estimation; Robust estimation;

Preliminary analysis; Shewhart chart.

INTRODUCTION assignable causes of variation. A slightly different

situation arises once the process is more or less in
Control charts are applied for a number of statistical control and the control chart is used to
purposes and in a range of contexts. Their original monitor the process. Literature on control charts
application is as part of a control loop, for instance, (Woodall and Montgomery, 1999) focuses predomi-
in the form of Shewhart’s continuing and self- nantly on the control chart in this context.
correcting operation to bring a process into a state Consequently, the current charting methodology
of statistical control (Shewhart, 1939, p. 25). In this has developed charts that function well in the context
application, measurements from the running process of monitoring, but this does not necessarily make
are added one by one to the chart. The chart’s them perform well as a technique in other applica-
purpose is to instigate and guide the search for tions. A different application of control charts is their

*Correspondence: Dr. Jeroen de Mast, Senior Consultant, Institute for Business and Industrial Statistics of the University
of Amsterdam (IBIS UvA), Plantage Muidergracht 24, 1018 TV, Amsterdam, The Netherlands; Fax: +31 20 5255101;
E-mail: [email protected].

407

DOI: 10.1081/QEN-120027943 0898-2112 (Print); 1532-4222 (Online)

Copyright & 2004 by Marcel Dekker, Inc. www.dekker.com
 ORDER REPRINTS

408 de Mast and Roes

use to judge from a given data set whether the process THE EXPLORATORY VS. THE
that produced the data is in statistical control. This MONITORING CONTEXT
type of application is called retrospective analysis
(Woodall and Montgomery, 1999). To illustrate the importance of studying control-
In this article we study the use of control charts charting procedures for exploratory analysis apart
for exploratory analysis. This type of analysis is from the monitoring context, we consider differences
performed on a given data set (such as in retro- between the two situations. These differences bear
spective analysis but unlike monitoring) and has as its upon:
purpose the detection of assignable causes of varia- 1. Continuing time series vs. finite sample. In the
tion, i.e., assist an inquirer who seeks to discover monitoring situation, samples are collected one by
sources of variation in a process by revealing evidence one. Each newly collected data point is compared to
of the variation sources. This purpose is identical to the control limits to verify whether the process is still
the function of control charting in the monitoring in statistical control. The probability
of a false
context but deviates from its application in retro- signal is associated with a tail area under the
spective analysis, where the (primary) purpose is not distribution of the individual measurements. If the
the detection of assignable causes (which is an normal distribution is taken as an approximation of
analysis of individual observations) but the inference this distribution, the control limits are usually set at a
whether or not the process is in statistical control distance of 3 of the central line, which corresponds
(which is an inference about the complete data with an approximate false alarm probability of
sample). Control charts are used in this manner in 0.0027 per observation.
exploratory studies (Hoaglin et al., 1983) as well as an In the exploratory case, the inquirer deals with a
after-the-trial check in the analysis of experimental finite sample. Consequently, he could consider basing
the control limits on an overall false alarm prob-
data to check for evidence of assignable causes. It
ability. The approximate false alarm probability
is
is the purpose of this article to develop a control-
then based on the joint distribution of all the data in
charting procedure that meets the needs of explora-
the sample. Woodall and Montgomery (1999) advo-
tory analysis.
cate this procedure in the context of retrospective
We limit ourselves to control charts for data that
analysis, and it seems appropriate when the chart
are collected as individual measurements, leaving
ought to provide answers relating to the complete
situations in which data are collected in subgroups
data set (such as, Was the process controlled?). We
for further research. Furthermore, we consider
shall argue that it is inappropriate in the exploratory
situations in which the data can be described reason-
context.
ably adequately by means of the normal distribution, 2. State of the process. Monitoring is especially
although the methodology is robustified against useful if the process has first been brought in a more
deviations from the normal distribution, especially or less controlled state. In the exploratory situation,
in the tails. Finally, we assume that the data are however, there is no reason to assume that the
independent. The control chart is designed with process is in control, hence it is very possible to
moderate sample sizes (say, 20 to 100 data points) encounter a chaotic sample of measurements. It is,
in mind. therefore, important to protect oneself by making the
The article is organized as follows. First we control chart robust against deviations from the
investigate the differences between the exploratory assumptions that underlie the method. In line with
context and the monitoring context. Two motivating exploratory data analysis theory in general, robust
examples are discussed, highlighting the problems methods ought to play a dominant role (Hoaglin
that are encountered if an inquirer applies control et al., 1983).
charts for exploratory analysis. The fourth section 3. Dependence between test statistics and control
enumerates the requirements for a control chart in limits. The control limits for a chart that is used for
the exploratory context and incorporates these in a monitoring are computed from an initial sample or
control-charting procedure. The statistical technical- from historical data. Hence, the measurements that
ities are developed in the fifth section. Two real-life are collected from the production process are
examples demonstrate the proposed procedure’s stochastically independent of the control limits. In
effectiveness in comparison with alternative con- the exploratory situation this is not the case: control
trol-charting methods. A discussion concludes the limits are computed from the measurements that are
article. plotted in the chart. As a consequence, disturbances
 ORDER REPRINTS

Robust Individuals Control Chart 409

6
4
2
0
-2
-4
-6

5 10 15 20 25

Figure 1. Control chart with control limits based on the average moving range.

in the observations can inflate the control limits, used for exploratory analysis is the individuals
making the chart less sensitive in detecting the control chart (Roes et al., 1993). In this control
remaining assignable causes. This is a second chart—referred to as the average moving range
motivation to use robust statistical procedures. (AMR) chart—the control limits are computed from:
4. Dependence between successive tests. If the
control limits are computed from the measurements UCL ¼ ^ þ 2:66 AMR
ð1Þ
that are indicated in the chart, the successive LCL ¼ ^
2:66 AMR
comparisons of observations to the control limits
are no longer independent, as was observed by Here, ^ is the overall mean of the data and:
Quesenberry (1993). Consequently, the fraction of
points that are mistakenly identified as outliers is no 1 X n
1
AMR ¼ jyi
yiþ1 j ð2Þ
longer
. Quesenberry shows that the dependence n
1 i¼1
between the successive tests is related to the variance
of the control limits. This is a motivation to base the The outliers in the data result in an inflated error
control limits on efficient estimators. estimate; consequently, the control limits are too
5. Difference in functionality. In monitoring, the wide: the chart in Fig. 1 gives one signal (observation
emphasis is on the signal function of the control 14)—and even this point is only just below the lower
chart, whereas the emphasis in exploratory analysis is control limit.
on clue generation. For this reason, the requirement The data in Fig. 2 were obtained from a
that the chart provides an easily interpreted display simulation of the model:
of the data is even more important in the exploratory yi ¼ i þ i ð3Þ
context than in the monitoring context.
with i ¼ 10.0 for i ¼ 1, . . . , 20; i ¼ 12.0 for
i ¼ 21, . . . , 30; and i ¼ 8.0 for i ¼ 31, . . . , 50. The i
TWO MOTIVATING EXAMPLES are i.i.d. N (0, 1). Observation 47 was set to the value
12.5. We would like a control chart to make the
Two examples demonstrate some of the problems following suggestions:
that could arise when control charts are applied in
exploratory analysis. . Shift in the mean at i ¼ 21 and i ¼ 31.
The data in Fig. 1 were drawn from the N (0, 1) . Outlier at i ¼ 47.
distribution. Observations 11 and 20 were replaced . The remaining variation is white noise.
with the value 6.0, observation 14 with the value

6.0. Given the way in which the data were obtained, The AMR chart gives one signal. It does not give
one would like a control chart to give three signals: at any information about the moment when the shifts
observations 11, 14, and 20. The chart that is often occurred, and the shifts make the detection of further
 ORDER REPRINTS

410 de Mast and Roes

14
12
10
8

0 10 20 30 40 50

Figure 2. Control chart with control limits based on the average moving range.

assignable causes—such as the outlier at i ¼ 47— randomness is an outlier. Other criteria should be
difficult. based on the order of data, in which generic patterns
could be observed by which assignable causes
manifest themselves. We propose to discern, apart
PROPOSED CONTROL CHART from outliers, one generic pattern that the control
chart should detect, namely shifts in the mean. The
Requirements importance of this pattern is acknowledged by the
extensive literature on CUSUM and EWMA charts.
In order to identify requirements for a control The discussion about the inclusion of more generic
chart for exploratory analysis, we study the require- patterns is presented in the last section of this article.
ments listed by Shewhart (Shewhart, 1939, p. 30) for As demonstrated in the examples in the previous
control charts: section, the presence of assignable causes of either
kind, outliers or shifts in the mean, seriously affects
1. They should indicate the presence of assign- the chart’s ability to detect the remaining assignable
able causes of variation. causes. Therefore, and in view of points 2 and 3 in
2. They should not only indicate the presence of section 2, the procedure should be robust against
assignable causes but also should do this in a both outliers and shifts in the mean. The former is
way to facilitate the discovery of these achieved by employing robust estimation methods,
causes. the latter by incorporating detected shifts in the
3. They should be as simple as possible and analysis. Finally, a clear visualization is vital in the
adaptable in a continuing and self-corrective exploratory context to facilitate the revelation of
operation of control. more nonrandom patterns, as noted in point 5 in
4. They should be such that the chance of section 2.
looking for assignable causes when they are
not present does not exceed some prescribed
value. Ad 2. Facilitate the Discovery of Detected
Assignable Causes
We discuss these requirements for the control
chart in the exploratory context. In exploratory analysis, this is an important
point and it is the reason why the chart should not
use an overall false alarm probability
. The inquirer
Ad 1. Indicate the Presence of Assignable Causes is interested in questions such as, when were assign-
able causes present? and how many assignable causes
Assignable causes are detected by anything that were present?, which are questions related to indivi-
indicates nonrandomness. An obvious deviation from dual measurements. Hence, the chart’s performance
 ORDER REPRINTS

Robust Individuals Control Chart 411

should be based on the distribution of the individual that the purpose of the control chart is not only to
data points and individual
s, such as in the specify when to search for assignable causes but also
monitoring context, are appropriate. when not to do this. In our experience, engineers find
it difficult to assess which patterns typically arise in
random noise and which are indicative for assignable
Ad 3. Simple and Adaptable causes. The control limits give the guidance that these
engineers need. This does not mean that the
The computations required for the set-up of a comparison of observations to control limits should
control chart are nowadays always performed by a be seen as a hypothesis test (Woodall, 2000). It only
computer, and as a consequence, the restrictions on means that our chart should indicate which points are
computational efforts are much looser than in suspicious and which are not and that our opera-
Shewhart’s days. We do not, therefore, consider tional definition of ‘suspicious’ is based on tail areas
simplicity or adaptability of computation an impor- of distribution functions.
tant requirement. Simplicity of the visual presenta-
tion, on the contrary, is important.

Proposed Procedure
Ad 4. Low False-Alarm Rate
For the analysis of individual measurements in
the exploratory context, we propose the subsequent
Shewhart warns against confusing criteria for the
identification of assignable causes and hypothesis control-charting procedure.
tests. This relates to the issue that as long as the
process is not in statistical control, the assumptions 1. Estimate the locations of possible shifts
that are required for a hypothesis test (e.g., concern- and test significance of these shifts. Upon
ing the assumed distribution) cannot be made completion of this step, the original data
rigorously. The mathematical theory of hypothesis set is divided into intervals on which the
testing serves as a background for the development mean of the measurements is presumed
of control-chart procedures, yet it is fundamentally constant.
different (Shewhart, 1939, pp. 39–40). In particular, 2. Estimate (using robust estimators) the means
the chance of looking for an assignable cause of the intervals between successive shifts.
when there is none present is associated with the Also, estimate (using robust estimators) the
tail probability of an assumed distribution function, variance of the in-control measurements.
but this association should be seen as a rough 3. Based on these estimates, determine a pair of
approximation. control limits for each interval. Measure-
Various authors have argued against plotting ments are identified as outliers if they fall
control limits in a control chart when it is used in the beyond these control limits.
context of exploratory data analysis. In reply to
Woodall (2000), Hoerl says that ‘‘the plot of the data A preview of the effectiveness of this procedure
over time is much more valuable than the control in dealing with the problems illustrated in section 3 is
limits.’’ However, Shewhart’s requirement 4 suggests shown in Fig. 3.
6

14
4

12
2
0

10
-2

8
-4

6
-6

5 10 15 20 25 0 10 20 30 40 50

Figure 3. Proposed control chart.

 ORDER REPRINTS

412 de Mast and Roes

Relationship with Propositions has nonetheless an efficiency that is only slightly less
in Literature than that of the sample standard deviation.
Robust methods for the individuals chart
Hawkins (1993) distinguishes isolated assignable were studied by Bryce et al. (1997) and Boyles
causes from persistent assignable causes and links (1997). The median absolute deviation (MAD) from
these to the Shewhart-type control chart for the first the median seems to outperform other estimators in
type and charts that accumulate information from situations in which outliers are present. However,
successive measurements (CUSUM, EWMA) for the neither of the two articles draws the biweight
second. The proposed control chart is a combination A-estimator or other advanced robust estimators
of both types of techniques: it employs change-point into the comparison.
analysis techniques to detect shifts and Shewhart-type In the following section we derive the necessary
control charts for the intervals between shifts. This estimation and testing procedures. In order to satisfy
combination has been proposed in the literature the robustness requirement, the theory of M-estima-
(Lucas, 1982; Sullivan and Woodall, 1996) but not in tors plays an important role. The reader is referred to
a form that has the two analyses incorporated in a Hoaglin et al. (1983) for a concise introduction.
single graph. There is a lot of literature on the
application of change-point analysis in control-
charting methodology, such as literature on ESTIMATION AND
CUSUM techniques (Page, 1954) and Cuscore TESTING PROCEDURES
charts (Box and Ramirez, 1992). Especially in the
context of retrospective analysis, a powerful method Estimation of the Location of a Shift
is proposed by Koning and Does (2000). However,
these articles focus on detection of shifts and do not Suppose that an inquirer has collected a sample of
address the issue of estimating their location. Sullivan n measurements that are to be analyzed using a control
and Woodall (1996) do consider estimation of the chart. The null model that is assumed for the in-
location of a shift. Their procedure is less powerful in control process has all measurements yi, i ¼ 1, . . . , n,
detecting shifts than the procedure of Koning and share the same normal distribution:
Does, but the gain in power of the latter is
yi ¼  þ i , i ¼ 1, . . . , n, ð4Þ
appearance, since the extra detected shifts are so
small that they cannot be estimated accurately. with i i.i.d. N (0,  2).
The role that robust estimation ought to play in To derive the required estimation and testing
control-charting methodology was well stated by methods, we start studying how to estimate the
Rocke (1989), who observed that: location  of a shift, given that a single shift has
occurred. The situation is described by the following
. statistics that are indicated in the control model:
chart should be sensitive to outliers, whereas

yi ¼ 1 þ i for i ¼ 1, . . . ,
. statistics that are used to calculate the control ð5Þ
yi ¼ 2 þ i for i ¼  þ 1, . . . , n
limits should be robust against outliers.
i i.i.d. N (0,  2); 1, 2, , and  are unknown.
In past decades, a number of articles has Sullivan and Woodall (1996) derive the maximum
appeared that deal with robust control charts. A likelihood (ML) estimator for . It is found, together
robust mean and range chart was proposed by with the ML estimators for 1, 2, and , from:
Langenberg and Iglewicz (1986), who base their ð~ 1 , ~ 2 , ~ , ~Þ ¼ arg min ‘ð1 , 2 , , Þ ð6Þ
control limits on the trimmed mean of the subgroup 1 , 2 , , 
means and the trimmed mean of ranges. Rocke with
(1989) studied the use of the inter quartile range
n
(IQR) as a robust estimator of error and found it to ‘ð1 , 2 , , Þ ¼ logð2 2 Þ
perform well in a number of out-of-control scenarios. 2
 
More advanced estimators for the mean and range 1 X
þ 2 ð yi
 1 Þ 2
chart where studied by Tatum (1997). He recom- 2 i¼1
mended the use of a variant of the biweight Xn 
2
A-estimator (Lax, 1985). This estimator is very robust þ ð y i
2 Þ ð7Þ
against deviations from the model assumptions but i¼þ1
 ORDER REPRINTS

Robust Individuals Control Chart 413

0
Arg min f (x) denotes the value of x for which f (x) is denoting the derivative of . The given procedure
minimized. Taking derivatives for 1, 2, and  and to obtain a scale estimate uses the asymptotic
equating to zero, we find for fixed : variance of the M-estimators for location in order
to estimate the standard deviation of the error. This
1X  eq. 12
type of scale estimator is named A-estimators by Lax
c~ 1 ½ ¼ yi
 i¼1 (1985). Lax finds A-estimators to perform well when
1 X n compared to other robust scale estimators, including
~ 2 ½ ¼ yi M-estimators. See also Hoaglin et al. (1983), Gross
n
 i¼þ1
! (1976), and Shoemaker (1984).
1 X Xn We propose to take for the derivative of the
2
~ ½ ¼ ðyi
~ 1 ½Þ2 þ ðyi
~ 2 ½Þ2 ð8Þ bisquare function:
n i¼1 i¼þ1


 ¼ 2, 3, . . . , n
2. Next, uð1
u2 Þ2 , juj  1,
ðuÞ ¼ ð13Þ
0, juj > 1:
~ ¼ arg min ‘ð~ 1 ½, ~ 2 ½, ~ ½, Þ
2n
2

¼ arg minflogð~ ½Þg ð9Þ Although this choice can be well motivated, we
mention Huber’s, Hampel’s and Andrews’ func-
Finally, tions as alternative options (Hoaglin et al., 1983).
Observations y1 , . . . , y in the neighborhood of the
~ 1 ¼ ~ 1 ½~; ~ 2 ¼ ~ 2 ½~; ~ ¼ ~ ½~ ð10Þ estimated mean ^ 1 ½ have standardized values
However, we are well aware that not all u ¼ ðyi
^ 1 ½Þ=cs0 ½ around 0, as do observations
measurements are necessarily produced by the yþ1 , . . . , yn in the neighborhood of ^ 2 ½. For these
in-control model Eq. (5). We anticipate that some observations, (u) behaves as ku (for a nonzero
observations—outliers—might stem from a more constant k), which is the -function associated with
heavily tailed distribution. In order to make the the ML estimator for normally distributed measure-
procedure robust against the possible presence ments. In view of the assumed normality of the center
of outliers, it is modified in the following way. of the distribution, this is an important property
(Hoaglin et al., 1983, p. 363). Observations further
P  ¼ 2, . . . , n
2, let ^ 1 ½ be the solution of
For
removed from the mean are more and more down-
P1 Þ=cs0 ½Þ ¼ 0 and let ^ 2 ½ be the
i¼1 ðð yi

weighted, and observations further removed than
solution of ni¼þ1 ðð yi
2 Þ=cs0 ½Þ ¼ 0, where:
cs0[] are completely rejected. Simulations show that
s0 ½ ¼ medianfjyi
mjgi¼1,...,n ð11Þ the value c ¼ 9 for the tuning constant results in a
good performance in a range of situations.
with m ¼ m1[] if 1  i  , and m ¼ m2[] if Analogous to Eq. (9), ^ could be determined
 þ 1  i  n, where m1[] and m2[] are the medians from:
of y 1, . . . , y and yþ1, . . . , yn, respectively.
Furthermore, is an odd function, whereas c is
a tuning constant. Thus, ^ 1 ½ and ^ 2 ½ are ^ ¼ arg minflogð^ ½Þg ð14Þ
M-estimators for location based on a preliminary
estimate s0[] for scale. Our final scale estimate is but because of the chosen function, evidence of
given by: (larger) shifts is mistakenly downweighted (or
rejected) as stemming from outliers. Therefore, a
^ ½
 better estimate for  is found from:
pffiffiffi P 2
ncs0 ½ i¼1 ðð yi
^ 1 ½Þ=cs0 ½Þ
1=2 ^ ¼ arg minflogð^ # ½Þg ð15Þ
Pn 2
þ i¼þ1 ð ðð yi
^ 2 ½Þ=cs0 ½Þ
¼  ð12Þ where ^ # ½ is found by substituting # for in Eq.
 P
 0
(12). Function # is a stretched version of so as to
 i¼1 ðð yi
^ 1 ½Þ=cs0 ½Þ
 reflect the fact that we have data from two normal
P 
þ ni¼þ1 0 ðð yi
^ 2 ½Þ=cs0 ½Þ distributions (with mean 1 and mean 2). Figure 4
illustrates the idea. Letting
¼ j^ 1 ½
^ 2 ½j=cs0 ½,
 ORDER REPRINTS

414 de Mast and Roes

-D 0 D -D 0 D
#
Figure 4. (left) and (right). The locations of u ¼ 0 and u ¼ 
are marked with 0 and D or
D.

we define:
Table 1. Suitable values for n1
#
ðuÞ and n2 for various sample sizes n.
8
>
> uð1
u2 Þ2 , juj  p1ffiffi5 , n n1 n2
>
>
>
>
< sgnðuÞ 25pffiffi5 , p1ffiffi <juj 
þ p1ffiffi ,
> 16 2.09 1.15
5 5
5
¼ sgnðuÞðjuj

Þ ð16Þ 8 2.57 1.95
>
> 10 2.98 3.00
>
>
>
> ð1
ðjuj

Þ2 Þ2 ,
þ p1ffiffi5 <juj 
þ 1, 15 3.26 5.70
>
:
0, juj>
þ 1: 20 3.50 10.90
30 3.76 29.60
#
We use only in Eq. (15); all other estimators are 40 3.97 55.30
based on . As in Eq. (10), ^ 1 ¼ ^ 1 ½^; ^ 2 ¼ ^ 2 ½^; 50 4.13 90.60
^ ¼ ^ ½^. 60 4.23 1
70 4.33 1
100 4.42 1
150 4.56 1
Testing Significance of a Shift

We test the null-model Eq. (4) against the

alternative model Eq. (5). The likelihood ratio
statisticPfor this testing problemPis ð~ 02 =~ 2 Þn=2 , with denominator. For various n, suitable choices for n1
n~ 02 ¼ ni¼1 ðyi
~ 0 Þ2 and n~ 0 ¼ ni¼1 yi . This statis- and n2 are given in Table 1. The appendix provides
tic can be shown to be equivalent with: details about the determination of these values.
rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi  The proposed approximation is accurate up to
~ðn
~Þ ~ 2
~ 1 the 0.99 quantile of the distribution of RT2/n1. An
T¼ ð17Þ
n  accurate approximation to the values for n1 and n2
can be found from the following formulas:
with  2 ¼ n~ 2 =ðn
2Þ. This statistic has the same
appearance as the test statistic in the two sample
t-test, but because of the fact that ~ is a random n1 4:58
22:4=n þ 52:2=n2
variable, it does not follow a t-distribution under the
n2 2:41
0:424n þ 0:0438n2
null model.
The test statistic defined in Eq. (17) can be ðFor n > 50, take n2 ¼ 1:Þ ð19Þ
robustified by plugging in the robust estimators
proposed in the preceding section. Thus, we arrive at: Altogether, the test signals if ð1=n1 ÞRT 2 > Fn
1 
1 ,n2
rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi  ð1

Þ, with
the probability of falsely detecting a
^ðn
^Þ ^ 2
^ 1
RT ¼ ð18Þ shift. Table 2 illustrates the performance of the
n ^
procedure. For shifts of various magnitudes (in
Simulations show that the distribution of RT2/n1 multiples of ) and for various sample sizes n, the
under the null model can be well approximated by an table gives the percentage of detected shifts for shifts
F-distribution having n1 degrees of freedom in the that occur on one fourth of the data series and
numerator and n2 degrees of freedom in the halfway the data series.
 ORDER REPRINTS

Robust Individuals Control Chart 415

Table 2. Percentage of detected shifts depending on magnitude of the shift, sample size, and location of the shift.

Sample size n and location of shift 

10 20 40 80
Shift
(multiples of ) 3 5 5 10 10 20 20 40

1 6.0% 6.3% 21.9% 29.3% 47.0% 61.8% 83.4% 93.4%

1.5 10.5% 10.2% 49.3% 62.4% 84.8% 95.5% 99.7% 100.0%
2 14.4% 19.6% 74.5% 88.7% 98.3% 99.8% 100.0% 100.0%
2.5 23.4% 28.7% 90.9% 98.2% 100.0% 100.0% 100.0% 100.0%
3 30.6% 38.9% 98.2% 99.8% 100.0% 100.0% 100.0% 100.0%

Multiple Shifts with mpthe ffiffiffi median of the relevant subgroup. The
factorpffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
n in the numerator of Eq. (12) is replaced
Having detected a shift, the data are split into with n2 =ðn
kÞ in order to account for the loss of
two groups: y1 , . . . , y^ and y^þ1 , . . . , yn . The same degrees of freedom for the estimation of 1 , . . . ,k .
procedure is applied to both groups, verifying Note that ^ is calculated from instead of #; this
whether more shifts can be detected. This procedure ensures that our final scale estimate is robust against
is continued recursively until no more shifts are nuisance stemming from slight misestimation of .
detected or until the size of the groups becomes For each group j ¼ l , . . . , k, the chart has control
smaller than 4 or any other chosen minimum value. limits at;
In the situation of a larger number of alternating sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
shifts, this procedure suffers from a masking pro- ^jþ1
^j
1
blem. Since the control chart is designed with UCLj ¼ ^ j þ h ^
^jþ1
^j
moderate sample sizes in mind, the number of shifts sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
will not often be that large and, as a consequence, the ^jþ1
^j
1
adequacy of the proposed control chart is not LCLj ¼ ^ j
h ^ ð22Þ
^jþ1
^j
seriously affected. However, the construction of a
procedure that deals more effectively with multiple where we define ^1 ¼ 0 and ^kþ1 ¼ n for notational
shifts is an interesting topic for further research. ease. The scalar h determines how much evidence
we require before we are prepared to identify
an observation as an outlier. We work with
Detection of Outliers theffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
p traditional valueffi h ¼ 3. The factor
ð^jþ1
^j
1Þ=ð^jþ1
^j Þ stems from the depen-
Denoting by ^ 1 , . . . ,^ k the estimated means of dency between the yi and the control limits.
the groups in between the detected shifts ^2 , . . . ,^k , we
re-estimate the error from the formula:
^ TWO EXAMPLES

P^1 2
nðcs0 Þ i¼1 ðð yi
^ 1 Þ=cs0 Þ þ   
In two examples stemming from recent projects,
1=2 we compare the method with other control charts.
Pn 2
þ i¼^k
1 þ1 ðð yi
^ k Þ=cs0 Þ The first example stems from the food industry. The
¼ pffiffiffiffiffiffiffiffiffiffiffi ð20Þ principal stage in the production of decaffeinated
 P ^1
n
k i¼1 0
ðð yi
^ 1 Þ=cs0 Þ þ    coffee is an extraction process during which most of
 the caffeine is removed from the coffee beans. During
Pn 
þ i¼^k
1 þ1
0
ðð yi
^ k Þ=cs0 Þ the extraction, an auxiliary substance DCM is added
to the beans and removed after the process. The
As before, s0 is the MAD: percentage of DCM that is absorbed by the beans
and as a result cannot be removed is an important
s0 ¼ medianfyi
mg ð21Þ quality characteristic.
 ORDER REPRINTS

416 de Mast and Roes

4
3
DCM content
2
1
0

0 20 40 60 80

Figure 5. Proposed control chart for DCM percentage measurements.

A quality improvement project was started in Figure 6 shows the AMR chart for the DCM
order to bring the DCM percentages safely below a percentage measurements. This chart detects the
certain requirement. In an early stage in the project, larger outliers. It does not, however, provide indica-
DCM percentage measurements were collected from tions about the presence of the shifts. Moreover, two
96 batches of coffee beans. The robust exploratory of the smaller outliers are not detected. The standard
control chart of these data, which are multiplied by a deviation of the in-control process is estimated to be
scalar for reasons of confidentiality, is presented in 0.58, which seems too large. This large estimate can
Fig. 5. be explained by the fact that it includes the additional
The control chart reveals the presence of several variation that is caused by the shifts and by the fact
assignable causes: it provides evidence that in the that the average moving range is not as robust an
time period in which the 96 measurements were estimator as the A-estimator that was used in the
collected eight isolated disturbances occurred as well proposed procedure.
as two shifts in the mean. Removing the outliers and The AMR chart can be augmented with addi-
correcting the remaining measurements for the shifts, tional runs rules (Nelson, 1984). These extra tests
the inquirer is left with measurements that can be make the chart more powerful in detecting shifts and
described reasonably well by a normal distribution drifts. We apply two popular rules to the DCM
with standard deviation 0.35. Note that the identifi- percentage data:
cation of outliers is possible because the risk of a false
signal is specified per observation (cf. section 4.1, ad. 1. Signal when nine consecutive measurements
2). Had we chosen to work with an overall false alarm are on the same side of the center line.
risk, it would have been unclear what guidance the 2. Signal when four out of five measurements
control limits provide for the identification of are on the same side of and more than a
assignable causes. distance of 1 sigma from the center line.
The unstable state of this process is a serious
complication for experimentation. In the first The first rule signals at observations 9, 10, 11, 12,
instance, improvement efforts should focus on pre- 13, 60, 61, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89,
ventive actions against the disturbances. The control 90, 91, 92, 93, 94, 95, 96. The second rule signals at
chart in the figure gives important indications of their observations 4, 5, 6, 7, 8, 10, 11, 12, 13, 56, 57, 58, 60,
nature and the time instants on which they occurred. 89, 90, 91. Clearly, the chart finds evidence for the
A good strategy would be to discuss the control chart shifts. However, it does not provide clear information
with the operators and process engineers who work about their number or the time instants on which
with the process and, if possible, to consult log books they occur. Moreover, the detected shifts are not
in order to find more information about the process incorporated in the analysis, and, as a consequence,
conditions during those time when disturbances the chart is less sensitive in detecting the remaining
occurred. assignable causes. Also CUSUM and EWMA charts
 ORDER REPRINTS

Robust Individuals Control Chart 417

4
3
DCM content
2
1
0

0 20 40 60 80

Figure 6. AMR chart for DCM percentage measurements.

(with sensibly chosen parameters) detect the shifts but valuable suggestion but not specific enough to
fail to show to what extent the remaining variation is pinpoint the assignable cause.
white noise. If the inquirer is less interested in establishing
The second example demonstrates the perfor- whether additional between-groups variation is
mance of our procedure in the case of a smaller data present, but rather wants to know whether this
set (30 observations). The example concerns the additional variation is random, he could plot an
lengths of biscuits when they leave the oven. In AMR-chart with the average lengths considered as
order to identify assignable causes of variation in individual observations (this procedure is advocated
their length, three biscuits were measured every 2 for the monitoring context in Does et al. (1999)). The
minutes, for 60 minutes on a row. The resulting data chart is shown in Fig. 9. The chart fails to detect
set does not consist of individual measurements. the shift in the mean (even when it is augmented with
However, the proposed control chart can be applied the previously mentioned runs rules). It is concluded
to the averages of the groups of three measurements. that the proposed control chart provides the most
The chart for these averages is presented in Fig. 7. revealing presentation of this data set.
The chart suggests that the average length has
shifted between minute 22 and minute 24. Inquiry
with the operators on duty revealed that the speed of
DISCUSSION
a conveyor belt was modified at (precisely) that
moment in order to adjust the length a bit. In
The term control has little bearing on the
addition to this adjustment, the average length
application of control charts in exploratory analysis.
appears in statistical control, indicating that no Consequently, the terms control chart and control
assignable causes should be sought on the basis of limits are somewhat misleading, and process behavior
these measurements. chart and natural process limits might be better terms
The traditional chart for subgrouped measure- (this terminology is used by Wheeler and Poling,
ments is the ðX ,RÞ control chart. The X -chart is 1998).
shown in Fig. 8. The control limits are computed It has been argued that charts for monitoring in
from an estimate of the within-groups spread based which the control limits are computed from small
on the average range (see formula 6–5 in samples are misleading (Quesenberry, 1993).
Montgomery (1991), p. 204). The chart gives four Simulations show that the probability that the
signals. The X -chart signals when there is (a proposed exploratory chart gives a misleading
significant amount of ) between-groups variation image is modest, even for smaller sample sizes.
that cannot be accounted for by the within-groups Judging a chart to be misleading either if it
variation alone. In view of this behavior, the four incorrectly detects a shift or if it gives two or more
signals would suggest that the length of the biscuits is false alarms for outliers, we find that for a sample size
not perfectly stable in time, which is in itself a of 40, the probability that the chart is misleading
 ORDER REPRINTS

418 de Mast and Roes

56.0
55.5
Length (mm.)
55.0
54.5

0 10 20 30 40 50 60
Minutes

Figure 7. Proposed control chart for the averages of lengths.

55.6
55.4
Length (mm.)
55.2
55.0
54.8
54.6

0 10 20 30 50 60 70
Minutes

Figure 8. X -chart for the lengths of biscuits.

is 5.72%. Simulations also confirm the high efficiency properties of procedures like this, however, are only
of the A-estimator: 84% for a sample size of 40 mediocre to bad (Hampel et al., 1986, pp. 56–71).
(compared to the standard deviation). The efficiency There is, moreover, an inherent problem in the given
of a scale estimator based on the AMR is 61%. procedure: outliers might inflate the error estimate so
An alternative for using robust estimators for the much that the chart becomes too insensitive to detect
mean and in-control standard deviation is the even a single outlier (the chart in Fig. 1 comes quite
following iterative procedure: close to this situation). The given iterative procedure
is inferior to the use of the A-estimator in the
1. Plot a provisional control chart of the data. proposed control chart.
2. Identify outliers. The proposed procedure could, and perhaps
3. Recalculate the control limits, ignoring the should, be generalized to include more generic
identified outliers. patterns. Obvious candidates include linear (or
polynomial) trends, autoregressive or moving average
Steps 2 and 3 could be done once or iteratively terms, and shifts in the variance. The selected generic
until no more outliers are detected. It should patterns should span most of the space of patterns in
be understood that this procedure is nothing but which assignable causes manifest themselves. The
a robust estimation procedure. The statistical procedure should suggest one or a number of possible
 ORDER REPRINTS

Robust Individuals Control Chart 419

56.0
55.5
Length (mm.)
55.0
54.5

0 10 20 30 40 50 60
Minutes

Figure 9. AMR-chart for the averages of lengths.

interpretations of the data and it should indicate to The first two moments of the ECDF approx-
what extent the given interpretations succeed in imating the distribution of RT2/n1 can be determined
reducing the measurements to white noise. This analogously, and equating these with the first two
procedure appears to be a valuable aid for explora- moments of the Fn1,n2 distribution (Johnson and Kotz,
tory data analysis. The inquirer must, however, 1970), we find:
always look for clues that the automated procedure
M1 n2
does not detect. Moreover, the suggested interpreta- ¼
n1 n2
2
tions should be interpreted as hypotheses and further ð24Þ
testing should be required before the inquirer can M2 n22 ðn1 þ 2Þ
2
¼
arrive at a final interpretation of a set of data. n1 n1 ðn2
2Þðn2
4Þ
Solving n1 and n2 from these equations gives:
4M2
2M12
APPENDIX n2 ¼ ð25Þ
M2
M12
2M1

Approximating the Distribution of RT2 n2
2

n1 ¼ M 1 ð26Þ
n2
For suitable choices of n1 and n2, we approximate
the distribution of RT2/n1 [see Eq. (18)] under model
Eq. (4) by the Fn1,n2 distribution. This appendix shows
how suitable values for n1 and n2 were obtained for Further Specifics
various sample sizes n.
For fixed n, we draw 10,000 samples of size n For n > 50, we find n2 to increase rapidly. Since
from a standard normal distribution and we calculate Fn1,n2 converges to X 2n1 =n1 if n2 ! 1 (for fixed n1 and
the corresponding 10,000 realizations of RT2. The X 2n1 =n1 being the distribution of a statistic having a
first two moments of the empirical cumulative chi-square distribution with n1 degrees of freedom,
distribution function (ECDF) that is induced by divided by n1), we approximate the distribution of
these 10,000 values are given by: RT2 by the X 2n1 distribution for sample sizes larger
than 50. Following the same idea as above, n1 is
X
1 10,000
M1 ¼ RTi2 determined from:
n i¼1
ð23Þ n1 ¼ M 1 ð27Þ
X
1 10,000
M2 ¼ RTi4 For n  20, we run into the problem that the
n i¼1
proposed approximation is inadequate in the remote
 ORDER REPRINTS

420 de Mast and Roes

99 oooo 99 oo
ooooo
oooo ooooo
oo
oooooo o
oo
oo
o
ooooo oooooo
oooo oooooo
oooooo
o oo
oooooo
o
o o
ooooo ooo
oooo
oooo oooooo
ooooo 95
oooooo
oooo
o oo
o
o ooo
ooooo ooooo
ooooo 90 ooooo
ooooo
95
ooooo
o
o o
oo
o
oo oo
oooo ooooo
ooooo ooooo
90 ooooo
o
o 80
o
oo
ooooo
oo oo
ooooo oooo
ooooo oooo
80 o
oo
ooooo o
ooooo
o
ooooo ooo
oooo
oooo 50
oooo
oooooo
o
o ooo
o
o
o
oo o
oo
50 oooo oooo
oooo oooo
o
oo
ooooo 10 o
oo
ooooo
10 oooo ooooo
ooooo ooooo

0 1 2 3 4 5 6 0 2 4 6 8 10 12 14

Figure 10. Probability plot for n ¼ 20 (left) and n ¼ 60 (right).

tails (beyond the 0.99 quantile); the remote tails of on the y-axis are Fn1,n2 (y) instead of y. The plots are
the distribution of RT2/n1 are heavier than those of representative of the plots for other values of n and
the approximating F-distribution. These heavy tails they show that the approximation is accurate up to
have a large impact on the moments of the ECDF; the 0.97 quantile and fairly accurate up to the 0.99
consequently, the found values for n1 and n2 do not quantile.
give the best approximation on the relevant domain,
namely up to the 0.99 quantile. Ideally, we would
approximate the ECDF truncated at the 0.99 quantile
by an F-distribution truncated at the 0.99 quantile. ABOUT THE AUTHORS
We would then equate moments of these truncated
distributions. Unfortunately, it seems not feasible to
obtain expressions for the moments of the truncated Jeroen de Mast works as a statistical consultant
F-distribution. and researcher for the Institute for Business and
Another approach to circumvent the problem is Industrial Statistics of the University of Amsterdam
2
inspired by the following idea. Writing RTðiÞ for the (IBIS UvA). He teaches courses in six sigma and
2
ordered RTi , we note that the points supervises improvement projects in Dutch industry.
2
ðRTðiÞ =n1 , Fn
1 ð i
1=2 ÞÞ, i ¼ 1, . . . , 9900 would lie on
1,n2 10,000 He wrote a Ph.D. thesis titled ‘‘Quality Improvement
the line y ¼ x if the approximation were perfect (up to from the Viewpoint of Statistical Method.’’ He is a
the 0.99 quantile). Using this idea; member of the executive committee of the European
9X900    Network for Business and Industrial Statistics
2
1 i
1=2 2
n1 ,n2 ¼arg min RTðiÞ =1
F1 ,2 ð28Þ (ENBIS).
1 ,2
i¼1
10,000 Kit C. B. Roes has worked for various Dutch
companies as a statistical consultant. He has been
i.e., n1 and n2 are chosen so that they minimize the L2
head of the Centre for Biometry Wageningen in the
distance from the line y ¼ x. In practice, n1 and n2
Netherlands. Currently, he is director of Biometrics
were found taking the values of Eq. (25) and Eq. (26)
of NV Organon.
as starting points for a numerical routine minimizing
Eq. (28).
The reported values in Table 1 were determined
from Eq. (28) (for n ¼ 5, 6, 8, 10, 15, and 20), Eq. (25),
and Eq. (26) (for n ¼ 30, 40, and 50), or Eq. (27) (for
n ¼ 60, 70, 100, and 150). The adequacy of the REFERENCES
proposed approximation for n ¼ 20 and n ¼ 60 is
demonstrated in Fig 10, which shows the probability Box, G. E. P., Ramirez, J. (1992). Cumulative score
plots (up to the 0.99 quantile) for the ECDF. It charts. Quality and Reliability Engineering
2
displays the points ðRTðiÞ =n1 ,Fn
1 ð i
1=2 ÞÞ. The indices
1 ,n2 10,000
International 8(1):17–27.
 ORDER REPRINTS

Robust Individuals Control Chart 421

Boyles, R. A. (1997). Estimating common-cause Montgomery, D. C. (1991). Statistical Quality

sigma in the presence of special causes. Control. 2nd ed. New York: Wiley.
Journal of Quality Technology 29(4):381–395. Nelson, L. S. (1984). The Shewhart control Chart—
Bryce, G. R., Gaudard, M. A., Joiner, B. L. (1997). tests for special causes. Journal of Quality
Estimating the standard deviation for indivi- Technology 16(4):237–239.
duals control charts. Quality Engineering Page, E. S. (1954). Continuous inspection schemes.
10(2):331–341. Biometrika 41:100–115.
Does, R. J. M. M., Roes, K. C. B., Trip, A. (1999). Quesenberry, C. P. (1993). The effect of sample size
Statistical Process Control in Industry. on estimated limits for X and X control charts.
Dordrecht: Kluwer. Journal of Quality Technology 25(4):237–247.
Gross, A. M. (1976). Confidence interval robustness Rocke, D. M. (1989). Robust control charts.
with long-tailed symmetric distributions. Technometrics 31(2):173–184.
Journal of the American Statistical Association Roes, K. C. B., Does, R. J. M. M., Schurink, Y.
71(354):409–416. (1993). Shewhart-type control charts for
Hampel, F. R., Ronchetti, E. M., Rousseeuw, P. J., individual observations. Journal of Quality
Stahel, W. A. (1986). Robust Statistics—The Technology 25(3):188–198.
Approach Based on Influence Functions. Shewhart, W. A. (1939). Statistical Method from the
Viewpoint of Quality Control. Washington,
New York: Wiley.
D.C.: The Graduate School of the
Hawkins, D. M. (1993). Robustification of cumula-
Department of Agriculture. [Reprinted by
tive sum charts by winsorization. Journal of
Dover Publications: New York, 1986].
Quality Technology 25(4):248–261.
Shoemaker, L. H. (1984). Robustness properties for a
Hoaglin, D. C., Mosteller, F., Tukey, J. W. (1983).
class of scale estimators. Communications in
Understanding Robust and Exploratory Data
Statistics A: Theory and Methods 13(1):15–28.
Analysis. New York: Wiley.
Sullivan, J. H., Woodall, W. H. (1996). A control
Johnson, N. I., Kotz, S. (1970). Continuous
chart for preliminary analysis of individual
Univariate Distributions—2. Boston: Houghton
observations. Journal of Quality Technology
Mifflin.
28(3):265–278.
Koning, A. J., Does, R. J. M. M. (2000). CUSUM
Tatum, L. G. (1997). Robust estimation of the
charts for preliminary analysis of individual process standard deviation for control charts.
observations. Journal of Quality Technology Technometrics 39(2):127–141.
32(2):122–132. Wheeler, D. J., Poling, S. R. (1998). Building
Langenberg, P., Iglewicz, B. (1986). Trimmed mean Continual Improvement—A Guide for Business.
X and R charts. Journal of Quality Technology Knoxville: SPC Press.
18(2):152–161. Woodall, W. H. (2000). Controversies and contra-
Lax, D. A. (1985). Robust estimators of scale: finite- dictions in statistical process control, with
sample performance in long-tailed symmetric discussion. Journal of Quality Technology
distributions. Journal of the American Statistical 32(4):341–378.
Association 80(391):736–741. Woodall, W. H., Montgomery, D. C. (1999).
Lucas, J. M. (1982). Combined Shewhart-CUSUM Research issues and ideas in statistical process
quality control schemes. Journal of Quality control. Journal of Quality Technology
Technology 14(2):51–59. 31(4):377–386.
 Request Permission or Order Reprints Instantly!

Interested in copying and sharing this article? In most cases, U.S. Copyright
Law requires that you get permission from the article’s rightsholder before
using copyrighted content.

All information and materials found in this article, including but not limited
to text, trademarks, patents, logos, graphics and images (the "Materials"), are
the copyrighted works and other forms of intellectual property of Marcel
Dekker, Inc., or its licensors. All rights not expressly granted are reserved.

Get permission to lawfully reproduce and distribute the Materials or order

reprints quickly and painlessly. Simply click on the "Request Permission/
Order Reprints" link below and follow the instructions. Visit the
U.S. Copyright Office for information on Fair Use limitations of U.S.
copyright law. Please refer to The Association of American Publishers’
(AAP) website for guidelines on Fair Use in the Classroom.

The Materials are for your personal use only and cannot be reformatted,
reposted, resold or distributed by electronic means or otherwise without
permission from Marcel Dekker, Inc. Marcel Dekker, Inc. grants you the
limited right to display the Materials only on your personal computer or
personal wireless device, and to copy and download single copies of such
Materials provided that any copyright, trademark or other notice appearing
on such Materials is also retained by, displayed, copied or downloaded as
part of the Materials and is not removed or obscured, and provided you do
not edit, modify, alter or enhance the Materials. Please refer to our Website
User Agreement for more details.

Request Permission/Order Reprints

Reprints of this article can also be ordered at

http://www.dekker.com/servlet/product/DOI/101081QEN120027943