Quality Engineering, 25:211–224, 2013

Copyright # Taylor & Francis Group, LLC

ISSN: 0898-2112 print=1532-4222 online
DOI: 10.1080/08982112.2013.769057

Robust CUSUM Control Charting

Hafiz Zafar Nazir1,
Muhammad Riaz2,4, ABSTRACT Cumulative sum (CUSUM) control charts are very effective in
Ronald J. M. M. Does3, detecting special causes. In general, the underlying distribution is supposed
Nasir Abbas1,4 to be normal. In designing a CUSUM chart, it is important to know how the
1
Department of Statistics, chart will respond to disturbances of normality. The focus of this article is to
University of Sargodha,
control the location parameter using a CUSUM structure and the major
Sargodha, Pakistan
2 concern is to identify the CUSUM control charts that are of more practical
Department of Mathematics
and Statistics, King Fahad value under different normal, non-normal, contaminated normal, and special
University of Petroleum and cause contaminated parent scenarios. In this study, we propose and compare
Minerals, Dhahran, Saudi Arabia the performance of different CUSUM control charts for phase II monitoring of
3
Department of Quantitative location, based on mean, median, midrange, Hodges-Lehmann, and trimean
Economics, IBIS UvA, University statistics. The average run length is used as the performance measure of the
of Amsterdam, Amsterdam, The
CUSUM control charts.
Netherlands
4
Department of Statistics,
Quaid-i-Azam University, KEYWORDS average run length (ARL), contaminated environments, control
Islamabad, Pakistan charts, CUSUM statistic, fast initial response, normality, process location

INTRODUCTION
Statistical process control is a collection of different techniques that
help differentiate between the common cause and the special cause var-
iations in the response of a quality characteristic of interest in a process.
Out of these techniques, the control chart is the most important one. It is
used to monitor the parameters of a process, such as its location and
spread. Some major classifications of the control charts are variable and
attributes charts, univariate and multivariate charts, and memoryless charts
(Shewhart’s type) and memory charts (like cumulative sum [CUSUM] and
exponentially weighted moving range [EWMA]). The commonly used
Shewhart’s variable control charts are the mean (X), median, and
mid-range charts for monitoring the process location and the range (R),
standard deviation (S), and variance (S2) charts for monitoring the process
variability (cf. Montgomery 2009). The main deficiency of Shewhart-type
control charts is that they are less sensitive to small and moderate shifts
in the process parameter(s).
Another approach to address the detection of small shifts is to use
Address correspondence to Dr. memory control charts. CUSUM control charts proposed by Page (1954)
Muhammad Riaz, Assistant Professor,
Department of Mathematics and and EWMA control charts proposed by Roberts (1959) are two commonly
Statistics, King Fahad University of used memory-type control charts. These charts are designed such that they
Petroleum and Minerals, P.O. Box
1737, Dhahran 31261, Saudi Arabia. use the past information along with the current information, which makes
E-mail: [email protected] them very sensitive to small and moderate shifts in the process parameters.
211
 In order to obtain efficient control limits for phase performance. McDonald (1990) proposed the use
II monitoring, it is generally assumed that the process of the CUSUM chart that is based on a nonparametric
has a stable behavior when the data are collected statistic. He used the idea of ranking the observations
during phase I (cf. Vining 2009; Woodall and first and then using those ranks in the CUSUM struc-
Montgomery 1999). Most of the evaluations of ture. Hawkins (1993) proposed a robust CUSUM
existing control charts depend on the assumptions chart for individual observations based on Winsori-
of normality, no contaminations, no outliers, and zation. MacEachern et al. (2007) proposed a robust
no measurement errors in phase I for the quality CUSUM chart based on the likelihood of the variate
characteristic of interest. In case of violation of these and named their newly proposed chart the RLCU-
assumptions, the design structures of the charts SUM chart. Li et al. (2010) proposed a nonparametric
lose their performance ability and hence are of less CUSUM chart based on the Wilcoxon rank-sum test.
practical use. There are many practical situations Reynolds and Stoumbos (2010) considered the
where non-normality is more common (see, for robustness of the CUSUM chart for monitoring the
example, Janacek and Miekle 1997). One of the process location and dispersion simultaneously. Midi
solutions to deal with this is to use control charts and Shabbak (2011) studied the robust CUSUM con-
that are robust against violations of the basic trol charting for the multivariate case. Lee (2011) pro-
assumptions, like normality. posed the economic design of the CUSUM chart for
To date, numerous papers on robust control charts non-normally serially correlated data. S. F. Yang
have been published. Langenberg and Iglewicz and Cheng (2011) proposed a new nonparametric
(1986) suggested using the trimmed mean X and R CUSUM chart that is based on the sign test. They
charts. Rocke (1992) proposed the plotting of X studied the ARL performance of the proposed chart
and R charts with limits determined by the mean of for monitoring different location parameters. Simi-
the subgroup interquartile ranges and showed that larly, much work has been done in the direction of
this method resulted in easier detection of outliers robust EWMA control charting; for example, see
and greater sensitivity to other forms of out-of- Amin and Searcy (1991), S. F. Yang et al. (2011),
control behavior when outliers are present. Tatum and Graham et al. (2012).
(1997) suggested an interesting method for robust Most of the CUSUM control charting techniques
estimation of the process standard deviation for con- discussed above are based on first transforming the
trol charts. Moustafa and Mokhtar (1999) proposed a observed data into a nonparametric statistic and then
robust control chart for location that uses the applying the CUSUM chart on that transformed stat-
Hodges-Lehmann and the Shamos-Bickel-Lehmann istic. Unlike these approaches, L. Yang et al. (2010)
estimators as estimates of location and scale para- proposed the use of a robust location estimator
meters, respectively. Wu et al. (2002) studied the (i.e., the sample median) with the CUSUM control
median absolute deviations–based estimators and structure. Extending their approach, in this article
their application to the X chart. Moustafa (2009) we present a robust CUSUM chart that is based on
modified the Shewhart chart by introducing the five different estimators for monitoring the process
median as a robust estimator for location and absol- location of phase II samples. The performance of
ute deviations to median as a robust estimator for the CUSUM chart with different robust estimators is
dispersion. Recently, properties and effects of viola- studied in the presence of disturbances to normality,
tions of ideal assumptions (e.g., normality, outliers contaminations, outliers, and special causes in the
free environment, no special causes, etc.) on the process of interest. Before moving on toward the
control charts have been studied in detail by Riaz robust estimators, we provide the basic structure of
(2008) and Schoonhoven et al. (2011a, 2011b). the CUSUM chart in the next section.
Some authors have also discussed the robustness
of the CUSUM chart to situations where the underly- THE CLASSICAL MEAN CUSUM
ing assumptions are not fulfilled. Lucas and Crosier
(1982) studied the robustness of the standard CUSUM
CONTROL CHART
chart and proposed four methods to reduce the The mean CUSUM control chart proposed by Page
effect of outliers on the average run length (ARL) (1954) has become one of the most popular methods

H. Z. Nazir et al. 212

to monitor processes. For a two-sided CUSUM chart, data. The sample mean X is defined as
Pn
we plot the two statistics Ciþ and Ci
against a single Xj
X ¼ j¼1 n , is a linear function of data, and is
control limit H. These two plotting statistics are
defined as: widely used to estimate the population location
parameter. According to Rousseeuw (1991), the
þ
 sample mean X performs well under the normality
Ciþ ¼ max½
0; ðX i
l0 Þ
K þ Ci
1  
; ½1 assumption but it is highly affected due to
Ci
¼ max 0;
ðX i
l0 Þ
K þ Ci
1

non-normality, is sensitive to outliers, and has zero
where i is the subgroup number, X i is the mean of breakdown point, which means that even a single
study variable X, l0 is the target mean of the study inconsistent observation can change its value.
variable X, and K is the reference value of the . Median: The sample median X ~ is defined as the
CUSUM scheme (cf. Montgomery 2009). The starting middle-order statistic for odd sample sizes and
value for both plotting statistics is usually taken the average of the two middle-order statistics in
equal to zero; that is, C0þ ¼ C0
¼ 0. Next, we plot case of even sample sizes. The median is a robust
these two statistics together with the control limit estimator, so it is less affected by non-normality.
H. It is concluded that the process mean has moved Dixon and Massey (1969) showed that the
upward if Ciþ > H for any value of i, whereas the efficiency of the median with respect to the mean
process mean is said to be shifted downwards if decreases with an increase in sample size and the
Ci
> H for any value of i. Thus, the CUSUM chart efficiency approaches 0.637 for large sample sizes.
is defined by two parameters—that is K and H—that . Mid-range: The mid-range is defined as
X þX
have to be chosen carefully, because the statistical MR ¼ ð1Þ 2 ðnÞ , where X(1) and X(n) are the lowest
properties of the CUSUM chart are sensitive to these and highest order statistics in a random sample
parameters. These two parameters are used in the of size n. It is highly sensitive to outliers because
standardized manner (cf. Montgomery 2009) and its design structure is based on only extreme
are given as: values of data and therefore it has a zero break-
down point. Although its use is not very common
K ¼ krX ; H ¼ hrX ; ½2 due to its low efficiency for mesokurtic distribu-
tions, its design structure has the ability to perform
where rX ¼ prXffiffinffi, rX is the standard deviation of the
good in the case of small samples from platykurtic
study variable X and n is the sample size. The
distributions. Ferrell (1953) presented a compari-
CUSUM statistic in [1] is given for the mean of
son of both the efficiency of estimation and
samples.
the detection of disturbances by medians and
midranges. He indicated an advantage of these less
DESCRIPTION OF THE PROPOSED conventional measures in detecting outliers while
not losing much efficiency in detecting a change
ESTIMATORS AND THEIR
in the central value.
CORRESPONDING CUSUM CHARTS . Hodges-Lehmann estimator: This estimator
Let h be the population location parameter that is defined as the median of the pairwise Walsh
will be monitored through control charting and h^ averages; that is, HL ¼ median((Xj þ Xk)=2, 1  j 
be its estimator based on a subgroup of size n. There k  n). The main advantage of the HL estimator is
are many choices for h^ out of which we consider that it is robust against outliers in a sample. It
here the following: mean, median, mid-range, has a breakdown point of 0.29 (i.e., 29% is the
Hodges-Lehmann estimator, and trimean. least portion of data contamination needed to
Based on a random sample X1, X2, . . . , Xn of size derive the estimate beyond all bounds; cf.
n, these estimators are defined as follows: Hettmansperger and McKean 1998). If the underlying
distribution for the data is normal, then the asymp-
. Mean: This estimator is included in the study to totic relative efficiency (ARE) of the HL estimator
provide a basis for comparison, because it is the relative to the sample mean is 0.955; otherwise, it is
most efficient estimator for normally distributed often greater than unity (cf. Lehmann 1983).

213 CUSUM Control Charting

. Trimean: The TriMean (TM) of a sample is the in terms of their ARLs, which is the expected number
weighted average of the sample median and two of subgroups before a shift is detected. The in-
quartiles: TM ¼ Q1 þ2Q4 2 þQ3 (cf. Tukey 1977; Wang control ARL is denoted by ARL0 and an out-of-
et al. 2007). It also equals the average of the control ARL is denoted by ARL1.
  A code developed in R language is used to
median and the mid-hinge: TM ¼ 12 Q2 þ Q1 þQ 2
3
,
simulate ARL. For the said purpose, first we have
where Qp is the pth quartile of the sample (cf.
simulated the standard errors for all estimators; that
Weisberg 1992). Like the median and the
is, 5  105 samples of size n are generated and for
mid-hinge, but unlike the sample mean, TM is
each sample the value of estimator is determined.
a statistically resistant L-estimator (a linear combi-
The standard deviation of those 5  105 values of
nation of order statistics) having a breakdown
estimator is represented by rb. Then the phase II
point of 25%. According to Tukey (1977), using h
samples of size n from different environment
TM instead of the median gives a more useful
(explained later in this section) are generated under
assessment of the location parameter. According
an in-control situation and the proposed charts are
to Weisberg (1992), the ‘‘statistical resistance’’
applied on those samples until an out-of-control
benefit of TM as a measure of the center (of a dis-
signal is detected. The respective sample number
tribution) is that it combines the median’s empha-
(when the shift is detected by the chart) is noted,
sis on center values with the mid-hinge’s attention
which is the in-control run length. This process is
to the extremes. If the underlying distribution for
repeated 10,000 times so that we can average out
the data is normal, then the ARE of the TM relative
these run lengths to get the ARL0. Finally, a shift in
to the sample mean is 0.83 (cf. Wang et al. 2007).
the form of d is introduced in the process to evaluate
the ARL1 performance of the proposed charts, where
The performance of these five estimators for
d is the difference between l0 and l1, and l1 is the
phase I analysis could be seen in Schoonhoven
shifted mean. We have chosen to have ARL0 ffi 370.
et al. (2011a). In the current article, we study the
performance of these estimators with the CUSUM
structure for the phase II analysis. Following [1], Uncontaminated Normal
the hb-CUSUM control chart statistics Siþ and Si
can Environment
be represented as:
An uncontaminated normal distribution
  means
h   i 9 2
that all observations come from N l0 ; r0 . The ARLs
Siþ ¼ max 0; hbi
m0
Kb þ Ci
1
þ =
of the CUSUM charts using different estimators
h   h i ½3
Si
¼ max 0;
hbi
m0
Kb þ Ci
1

; under the uncontaminated environment are given
h
in Tables 1 and 2.
In [3], h^ may be any choice of the estimators men- Tables 1 and 2 indicate that all estimators have
tioned above. Initial values for the statistics given almost the same in-control ARL; that is, ARL0 ffi 370.
in [3] are taken equal to zero; that is, S0þ ¼ S0
¼ 0. For the out-of-control case, the ARLs of the CUSUM
The decision rule for the proposed chart(s), follow- based on the mean are clearly smaller compared to
ing the spirit of [2], is given in the form of Kh^ and Hh^ the other estimators. Here the performance of the
defined as: HL estimator is comparable to the mean estimator
because the ARL performance of both is almost the
Kh^ ¼ kh^rb; Hh^ ¼ hh^rb ½4 same, whereas the other estimators have larger
h h
ARL1 values. Moreover, the effect of change in
subgroup size and kh^ is identical for all estimators.
PERFORMANCE OF THE PROPOSED To further describe the run length distribution, we
also report the standard deviations of the run length
CUSUM CONTROL CHARTS (SDRL) and some percentile points of the run length
The performance of the proposed CUSUM control distribution in Tables 3 and 4, respectively.
charting design structures given in [3] and [4] is eval- Considering the results of Tables 3 and 4, we
uated in this section. The performance is measured observe that the run length distribution of all of

H. Z. Nazir et al. 214

 TABLE 1 ARL Values for the CUSUM Chart Based on Different Estimators under Uncontaminated Normal Distribution with
kh^ ¼ 0:25 and hh^ ¼ 8:03

Subgroup size Estimator 0 0.25 0.5 0.75 1 1.5 2

n¼5 Mean 370.096 24.592 10.006 6.353 4.645 3.151 2.395

Median 372.498 31.594 12.377 7.702 5.600 3.734 2.850
Mid-range 370.754 29.815 11.677 7.309 5.352 3.545 2.729
HL 373.121 25.834 10.438 6.545 4.806 3.245 2.476
TM 373.932 27.589 10.918 6.900 5.066 3.393 2.599
n ¼ 10 Mean 372.356 15.352 6.722 4.391 3.322 2.251 1.976
Median 372.746 18.954 8.053 5.194 3.877 2.653 2.057
Mid-range 373.508 23.407 9.578 6.043 4.503 3.039 2.292
HL 370.992 15.988 6.963 4.539 3.420 2.315 1.990
TM 375.278 16.773 7.284 4.707 3.526 2.411 2.001

the proposed CUSUM charts is positively skewed as and S0 ¼ Hh^=2 at ARL0 ffi 370. The results are given
long as there is some variation in the run lengths. in Table 5, where the uncontaminated normal distri-
By decreasing the value of kh^ from 0.5 to 0.25, bution is taken as the parent environment.
the standard deviation of all of the proposed charts It is gratifying to note that the FIR feature indeed
also decreases for small values of d. The percentiles enhances the performance of the proposed CUSUM
can also be used to compare the median run length charts. In the upcoming sections we will not give
(MRL) for CUSUM charts with different estimators. the additional tables for the SDRL, percentile points
Lucas and Crosier (1982) proposed the use of a of the run length, and the FIR features, although they
fast initial response (FIR) feature with the CUSUM can be easily obtained along the same lines.
charts in which they recommended not to set the
initial values of CUSUM statistics equal to zero. They Variance-Contaminated Normal
found that the choice of a head start S0 ¼ Hh^=2 is
optimal in the sense that its effect on the ARL0 is very
Environment
minor but it significantly decreases the ARL1 values. A (u) 100% variance-contaminated normal distri-
The intention of FIR was to enhance the CUSUM bution is one that contains
 (1
u) 100% observa-
2
chart sensitivity in detecting the shifts that occur tions from
 N l0 ; r0 and (u) 100% observations
immediately after the start of the process. As an from N l0 ; sr20 , where 0 < s < 1. The ARL values
example, we provide the ARL values of FIR CUSUM of the CUSUM charts using different estimators for
based on different estimators for n ¼ 10, kh^ ¼ 0:5, the variance-contaminated normal environment are

TABLE 2 ARL Values for the CUSUM Chart Based on Different Estimators under Uncontaminated Normal Distribution with
kh^ ¼ 0:5 and hh^ ¼ 4:774

Subgroup size Estimator 0 0.25 0.5 0.75 1 1.5 2

n¼5 Mean 370.469 28.338 8.304 4.788 3.383 2.242 1.803

Median 374.278 41.831 11.267 6.073 4.205 2.665 2.065
Mid-range 370.110 37.528 10.265 5.712 3.974 2.534 1.986
HL 367.095 29.990 8.790 4.999 3.521 2.306 1.851
TM 368.020 32.519 9.363 5.247 3.699 2.392 1.908
n ¼ 10 Mean 367.292 14.845 5.156 3.177 2.364 1.708 1.148
Median 367.663 19.538 6.390 3.837 2.774 1.946 1.468
Mid-range 368.529 26.582 7.919 4.568 3.275 2.175 1.756
HL 373.781 15.607 5.393 3.277 2.438 1.753 1.194
TM 370.949 16.720 5.640 3.421 2.528 1.816 1.261

215 CUSUM Control Charting

 TABLE 3 SDRL Values for the CUSUM Chart Based on Different Estimators under Uncontaminated Normal Distribution
with n ¼ 10 at ARL0 ffi 370

Chart parameters Estimator 0 0.25 0.5 0.75 1 1.5 2

kh^ ¼ 0:25 Mean 360.523 6.590 1.902 0.994 0.641 0.437 0.164
hh^ ¼ 8:03 Median 356.984 9.085 2.477 1.264 0.824 0.533 0.241
Mid-range 360.781 12.457 3.195 1.611 1.021 0.564 0.459
HL 350.709 6.980 1.991 1.032 0.669 0.469 0.142
TM 356.536 7.674 2.143 1.099 0.706 0.501 0.130
kh^ ¼ 0:5 Mean 367.992 9.498 1.963 0.939 0.577 0.465 0.355
hh^ ¼ 4:774 Median 363.205 13.603 2.766 1.245 0.764 0.403 0.499
Mid-range 361.144 20.159 3.764 1.623 0.975 0.479 0.446
HL 367.477 10.194 2.141 0.988 0.626 0.450 0.395
TM 365.022 11.056 2.278 1.043 0.647 0.421 0.439

given in Table 6 with / ¼ 0.05 and s ¼ 9 and in the most by this variance contamination. Similarly,
Table 7 with / ¼ 0.1 and s ¼ 9. in terms of ARL1 values, TM and HL outperform all
In Tables 6 and 7, the ARL0 for the TM and HL of the other estimators under discussion.
CUSUM charts are affected least by the contami-
nation, whereas the mid-range estimator is affected Location-Contaminated Normal
Environment
TABLE 4 Percentile Run Length Values for the CUSUM Chart
Based on Different Estimators under Uncontaminated Normal Dis- A (u) 100% location-contaminated normal distri-
bution is the one that  contains (1
u) 100% observa-
tribution with n ¼ 10, kh^ ¼ 0:5, and hh^ ¼ 4:774 at ARL0 ffi 370
2
d tions
 from N l r
0 0 and (u) 100% observations from
;
2
N l0 þ xr0 ; r0 , where
1 < x < 1. The ARL
Estimator Percentile 0 0.25 0.5 0.75 1 1.5 2
values of the CUSUM chart using different estimators
Mean P10 44 6 3 2 2 1 1 for a location-contaminated normal environment are
P25 111 8 4 3 2 1 1 given in Table 8 with / ¼ 0.05 and x ¼ 4.
P50 255 12 5 3 2 2 1
Table 8 shows that none of the estimators is able
P75 504 19 6 4 3 2 1
P90 833.1 27 8 4 3 2 2
to adequately detect the location contamination in
Median P10 45 7 3 2 2 1 1 the process when n ¼ 5, because the ARL0 for all
P25 112 10 4 3 2 2 1 the estimators is substantially lower than that for
P50 256 16 6 4 3 2 1 the uncontaminated environment. Increasing the
P75 501 25 8 5 3 2 2 subgroup size may be a better option because the
P90 843 37 10 5 4 2 2
mid-range and median CUSUM charts have a reason-
Mid-range P10 46 8 4 3 2 2 1
P25 113 12 5 3 3 2 1
able ARL0 for n ¼ 10 but the ARL1 performance of the
P50 255 21 7 4 3 2 2 mid-range CUSUM chart is way too poor compared
P75 504 35 10 5 4 2 2 to the median CUSUM.
P90 828 53 13 7 4 3 2
HL P10 44 6 3 2 2 1 1
P25 111 9 4 3 2 1 1
Special Cause Environment
P50 260 13 5 3 2 2 1 Asymmetric variance disturbances are created in
P75 518 20 7 4 3 2 1
which each observation is drawn from N(0,1) and
P90 847.1 29 8 5 3 2 2
TM P10 44 6 3 2 2 1 1
has a u probability of having a multiple of a v2ð1Þ vari-
P25 109 9 4 3 2 2 1 able added to it, with a multiplier equal to 4. ARLs of
P50 258 14 5 3 2 2 1 the CUSUM chart under this environment with
P75 517 21 7 4 3 2 2 / ¼ 0.01 and / ¼ 0.05 are given in Tables 9 and 10,
P90 846.1 31 9 5 3 2 2 respectively.

H. Z. Nazir et al. 216

 TABLE 5 ARL Values for the FIR CUSUM Chart Based on Different Estimators under Uncontaminated Normal Distribution
with n ¼ 10, kh^ ¼ 0:5, and S0 ¼ Hh^=2 at ARL0 ffi 370

Estimator hh^ 0 0.25 0.5 0.75 1 1.5 2

Mean 4.86 371.772 9.920 3.088 1.919 1.436 1.031 1.000

Median 4.87 371.999 14.114 3.886 2.297 1.694 1.134 1.007
Mid-range 4.9 370.930 19.588 4.853 2.699 1.969 1.295 1.043
HL 4.86 369.097 10.216 3.182 1.979 1.479 1.045 1.001
TM 4.86 372.281 11.417 3.364 2.069 1.523 1.068 1.002

In the presence of special causes, the median followed by the HL CUSUM. Median CUSUM has also
CUSUM seems more robust, whereas the mid-range reasonable performance compared to the others, but
CUSUM is affected the most. TM and HL CUSUM the mid-range CUSUM seems to have worst perfor-
have good detection ability with a reasonable ARL0. mance for the said case (cf. Table 11). For the logistic
TM and HL estimators are affected positively by the distribution, the HL and TM CUSUM outperform the
increase in subgroup size; that is, their ARL0 increase median and mid-range CUSUM charts, whereas the
as we increase the value of n and vice versa. mean CUSUM reasonably maintains its performance
(cf. Table 12). Similarly, TM and mean CUSUM charts
show very good performance in case of chi-square
Non-normal Environments
distribution. HL also performs well, whereas the
To investigate the effect of using non-normal distri- mid-range CUSUM has the worst performance (cf.
butions, we consider two cases: one by changing the Table 13).
kurtosis and the other by changing the symmetry of We provide the ARL curves of the CUSUM charts
the distribution. For the case of disturbing the kurtosis, with different estimators under different environ-
we use Student’s t distribution with 4 degrees of free- ments discussed above. Figures 1–5 contain the
dom (T4) and the logistic distribution (Logis(0,1)), and ARL curves of the CUSUM charts based on different
for the disturbance in symmetry we use the chi-square
  estimators with n ¼ 10, kh^ ¼ 0:5, and hh^ ¼ 4:774.
distribution with 5 degrees of freedom v2ð5Þ . From Figures 1–5 we see that the ARLs of the
Tables 11–13 contain the ARL values for the proposed mid-range CUSUM are affected the most in case of
CUSUM charts under T4, Logis(0,1), and v2ð5Þ , respect- some non-normal, contaminated, and special cause
ively, where the ARL0 is kept fixed at 370. environments. The mean CUSUM is also poorly
For the case of Student’s t distribution, TM CUSUM influenced under special cause environments. The
performs the best among all of the other estimators ARL0 values of the median, TM, and HL CUSUM

TABLE 6 ARL Values for the CUSUM Chart Based on Different Estimators under 5% Variance-Contaminated Normal Dis-
tribution with kh^ ¼ 0:5 and hh^ ¼ 4:774

Subgroup size Estimator 0 0.25 0.5 0.75 1 1.5 2

n¼5 Mean 303.375 40.319 10.911 5.900 4.129 2.630 2.037

Median 353.227 44.057 11.895 6.454 4.424 2.789 2.125
Mid-range 198.394 77.199 20.560 9.772 6.300 3.746 2.733
HL 342.091 35.703 10.112 5.604 3.918 2.511 1.975
TM 354.750 36.583 10.288 5.700 3.955 2.527 1.988
n ¼ 10 Mean 330.826 20.222 6.475 3.844 2.803 1.959 1.474
Median 366.873 21.549 6.797 4.009 2.891 1.995 1.545
Mid-range 197.230 83.124 22.377 10.551 6.622 3.920 2.866
HL 359.458 17.524 5.845 3.552 2.601 1.864 1.332
TM 361.461 18.050 6.040 3.606 2.652 1.886 1.350

217 CUSUM Control Charting

 TABLE 7 ARL Values for the CUSUM Chart Based on Different Estimators under 10% Variance-Contaminated Normal
Distribution with kh^ ¼ 0:5 and hh^ ¼ 4:774

Subgroup size Estimator 0 0.25 0.5 0.75 1 1.5 2

n¼5 Mean 296.883 51.619 13.442 7.054 4.816 2.984 2.226

Median 340.203 49.462 12.875 6.927 4.704 2.924 2.222
Mid-range 222.020 99.674 29.238 13.460 8.273 4.683 3.340
HL 306.853 42.497 11.465 6.248 4.288 2.723 2.077
TM 335.199 41.117 11.131 6.152 4.195 2.688 2.067
n ¼ 10 Mean 329.818 25.775 7.771 4.482 3.196 2.141 1.727
Median 356.416 22.899 7.161 4.213 3.026 2.057 1.624
Mid-range 242.112 113.064 34.908 15.300 9.298 5.152 3.642
HL 342.307 20.108 6.486 3.853 2.785 1.961 1.475
TM 361.443 20.064 6.473 3.807 2.774 1.950 1.464

TABLE 8 ARL Values for the CUSUM Chart Based on Different Estimators under 5% Location-Contaminated Normal
Distribution with kh^ ¼ 0:5 and hh^ ¼ 4:774

Subgroup size Estimator 0 0.25 0.5 0.75 1 1.5 2

n¼5 Mean 317.549 48.560 13.543 7.138 4.801 2.973 2.222

Median 301.845 49.447 13.245 6.928 4.682 2.933 2.210
Mid-range 315.737 78.997 25.153 12.361 7.818 4.429 3.124
HL 285.025 45.634 12.369 6.565 4.485 2.806 2.124
TM 297.391 43.218 11.775 6.330 4.372 2.742 2.085
n ¼ 10 Mean 344.932 24.938 7.639 4.450 3.155 2.106 1.710
Median 357.296 22.676 7.076 4.160 2.979 2.038 1.601
Mid-range 391.235 78.931 24.038 11.693 7.523 4.367 3.097
HL 303.981 21.554 6.811 3.978 2.879 1.986 1.543
TM 329.528 20.369 6.464 3.869 2.793 1.957 1.489

seem less affected by the change of the parent the charts under T4, Logis(0,1), and v2ð5Þ distributions
normal environment. are given in Figures 6, 7, and 8, respectively, with
For a graphical comparison of the proposed charts kh^ ¼ 0:5 and ARL0 fixed at 370. These figures clearly
with non-normal environments, the ARL curves of all indicate that, in general, TM and HL CUSUM are

TABLE 9 ARL Values for the CUSUM Chart Based on Different Estimators under Special Cause Normal Distribution with
u ¼ 0.01, kh^ ¼ 0:5, and hh^ ¼ 4:774

Subgroup size Estimator 0 0.25 0.5 0.75 1 1.5 2

n¼5 Mean 221.158 52.954 12.241 6.338 4.295 2.699 2.065

Median 365.619 41.318 11.252 6.094 4.230 2.671 2.052
Mid-range 168.113 132.188 48.782 15.514 8.731 4.740 3.301
HL 356.893 31.754 9.094 5.156 3.607 2.359 1.888
TM 364.754 33.872 9.704 5.368 3.762 2.418 1.921
n ¼ 10 Mean 211.164 23.559 6.853 3.989 2.861 1.987 1.552
Median 360.264 19.982 6.552 3.834 2.798 1.956 1.477
Mid-range 124.994 104.625 75.651 27.395 12.790 6.089 4.115
HL 367.158 15.936 5.460 3.345 2.465 1.784 1.215
TM 375.058 16.858 5.690 3.464 2.546 1.834 1.281

H. Z. Nazir et al. 218

 TABLE 10 ARL Values for the CUSUM Chart Based on Different Estimators under Special Cause Normal Distribution with
u ¼ 0.05, kh^ ¼ 0:5, and hh^ ¼ 4:774

Subgroup size Estimator 0 0.25 0.5 0.75 1 1.5 2

n¼5 Mean 132.301 80.360 32.039 13.488 7.879 4.404 3.095

Median 354.591 45.811 12.137 6.518 4.436 2.803 2.130
Mid-range 104.606 90.670 76.205 58.787 38.613 14.055 7.940
HL 292.820 45.662 11.647 6.180 4.230 2.680 2.056
TM 322.227 38.766 10.660 5.794 4.076 2.573 2.008
n ¼ 10 Mean 148.795 52.119 14.593 7.137 4.698 2.896 2.118
Median 368.855 21.345 6.730 3.980 2.866 1.994 1.534
Mid-range 105.884 91.084 75.093 62.220 48.336 24.478 12.724
HL 345.449 18.332 6.031 3.625 2.645 1.878 1.357
TM 355.257 18.601 6.020 3.639 2.656 1.889 1.366

TABLE 11 ARL Values for the CUSUM Chart Based on Different Estimators under T4 Distribution with n ¼ 10 and
kh^ ¼ 0:5 at ARL0 ffi 370

Estimator hh^ 0 0.25 0.5 0.75 1 1.5 2

Mean 4.99 371.243 30.619 8.690 4.933 3.511 2.309 1.877

Median 4.83 371.753 24.568 7.443 4.366 3.123 2.110 1.682
Mid-range 5.80 369.540 205.442 53.968 20.771 12.092 6.505 4.501
HL 4.846 371.437 22.654 7.019 4.138 2.986 2.043 1.607
TM 4.84 370.038 22.075 6.877 4.055 2.932 2.024 1.581

TABLE 12 ARL Values for the CUSUM Chart Based on Different Estimators under Standard Logistic Distribution
with n ¼ 10 and kh^ ¼ 0:5 at ARL0 ffi 370

Estimator hh^ 0 0.25 0.5 0.75 1 1.5 2

Mean 4.790 371.335 47.114 12.424 6.671 4.541 2.860 2.166

Median 4.817 369.302 52.499 13.795 7.246 4.944 3.075 2.304
Mid-range 4.970 370.869 126.070 35.936 15.877 9.760 5.415 3.808
HL 4.817 368.761 45.081 12.161 6.498 4.479 2.819 2.139
TM 4.813 371.226 45.926 12.258 6.630 4.493 2.817 2.150

TABLE 13 ARL Values for the CUSUM Chart Based on Different Estimators under Chi-square Distribution with n ¼ 10
and kh^ ¼ 0:5 at ARL0 ffi 370

Estimator hh^ 0 0.25 0.5 0.75 1 1.5 2

Mean 4.85 370.487 115.012 36.877 17.094 10.376 5.630 3.937

Median 4.90 369.970 127.209 44.319 20.670 12.017 6.527 4.454
Mid-range 5.19 369.765 196.527 96.916 49.213 28.913 13.261 8.395
HL 4.88 371.978 117.913 37.105 17.439 10.544 5.796 4.019
TM 4.87 370.087 115.104 37.648 17.085 10.494 5.760 4.000

219 CUSUM Control Charting

FIGURE 1 ARL curves of the mean CUSUM with n ¼ 5, kh^ ¼ 0:5, FIGURE 4 ARL curves of the HL CUSUM with n ¼ 5, kh^ ¼ 0:5,
and hh^ ¼ 4:774 under different parent environments. and hh^ ¼ 4:774 under different parent environments.

FIGURE 2 ARL curves of the median CUSUM with n ¼ 5, FIGURE 5 ARL curves of the TM CUSUM with n ¼ 5, kh^ ¼ 0:5,
kh^ ¼ 0:5, and hh^ ¼ 4:774 under different parent environments. and hh^ ¼ 4:774 under different parent environments.

performing better than the other estimators under chart (NPCUSUM) by S. F. Yang and Cheng (2011)
Student’s t and logistic distributions. All of the esti- under different environments discussed previously.
mators (except the mid-range) perform equally well S. F. Yang and Cheng (2011) calculated the ARLs of
in case of the chi-square parent environment. NPCUSUM using the shift parameter p1. For a valid
Finally, we provide a comparison of our proposed comparison of NPCUSUM with our proposed charts,
CUSUM charts with the nonparametric CUSUM mean we evaluated the ARL values of NPCUSUM chart

FIGURE 3 ARL curves of the midrange CUSUM with n ¼ 5, FIGURE 6 ARL curves of different CUSUM charts under T4
kh^ ¼ 0:5, and hh^ ¼ 4:774 under different parent environments. distribution with n ¼ 10, kh^ ¼ 0:5 and ARL0 ffi 370.

H. Z. Nazir et al. 220

FIGURE 7 ARL curves of different CUSUM charts under logistic
distribution with n ¼ 10, kh^ ¼ 0:5, and ARL0 ffi 370.

FIGURE 9 For uncontaminated data: (a) patients’ waiting

times; (b) output of mean CUSUM; (c) output of median CUSUM;
FIGURE 8 ARL curves of different CUSUM charts under (d) output of mid-range CUSUM; (e) output of HL CUSUM; and
chi-square distribution with n ¼ 10, kh^ ¼ 0:5, and ARL0 ffi 370. (f) output of TM CUSUM.

TABLE 14 ARL Comparison for the NPCUSUM (with n ¼ 10, k ¼ 0.5, and h ¼ 10.65) and TM CUSUM (with n ¼ 10 and kh^ ¼ 0:5) Charts
under Different Environments with Prefixed ARL0 ¼ 370

Environment Chart 0 0.25 0.5 0.75 1 1.5 2

Normal NPCUSUM 371.051 20.492 8.266 5.465 4.274 3.330 3.030

TM CUSUM 370.949 16.720 5.640 3.421 2.528 1.816 1.261
5% Variance contaminated NPCUSUM 370.868 21.764 8.626 5.618 4.422 3.444 3.082
TM CUSUM 361.461 18.050 6.040 3.606 2.652 1.886 1.350
10% Variance contaminated NPCUSUM 370.739 22.915 9.037 5.866 4.571 3.548 3.156
TM CUSUM 361.443 20.064 6.473 3.807 2.774 1.950 1.464
5% Location contaminated NPCUSUM 172.225 15.595 7.558 5.205 4.162 3.289 3.025
TM CUSUM 329.528 20.369 6.464 3.869 2.793 1.957 1.489
Special cause (u ¼ 0.01) NPCUSUM 367.505 19.645 8.157 5.407 4.253 3.331 3.029
TM CUSUM 375.058 16.858 5.690 3.464 2.546 1.834 1.281
Special cause (u ¼ 0.05) NPCUSUM 243.051 16.922 7.716 5.244 4.183 3.307 3.029
TM CUSUM 355.257 18.601 6.020 3.639 2.656 1.889 1.366
T4 NPCUSUM 370.492 22.744 9.075 5.940 4.685 3.682 3.263
TM CUSUM 370.038 22.075 6.877 4.055 2.932 2.024 1.581
Logis(0,1) NPCUSUM 371.779 43.808 15.060 8.956 6.624 4.604 3.814
TM CUSUM 371.226 45.926 12.258 6.630 4.493 2.817 2.150

221 CUSUM Control Charting

using the shift parameter as d, which is the difference these charts are given in Figures 9b–9f. For the 10%
between l0 and l1. The ARLs are calculated through variance contamination environment, we chose 10%
Monte Carlo simulations by performing 10,000 observations randomly from the data and inflated their
replications of run lengths and are given in variance nine times; that is, s ¼ 9. For 10% variance-
Table 14. For v2ð5Þ , the conversion from p1 to d is contaminated data, the plot of data and the output
not possible due to skewness of the distribution. of five charts is provided in Figure 10. Along similar
In some of the scenarios (such as location and 5% lines, the location contamination and the outliers are
special cause contaminations), the NPCUSUM will per- introduced into the data and the respective data plots
form better for small values of d ¼ 0.25, but, in general, and chart outputs are given in Figures 11 and 12.
the TM (and HL) CUSUM have better performance We can see that all of the proposed charts
across all the scenarios and for most values of d. show the stability for uncontaminated distribution
in Figure 9. For Figure 10, where the variance
A NUMERICAL EXAMPLE contamination is introduced into the process, the
mid-range CUSUM detects a positive shift at sample
In this section, we demonstrate the practical 14, whereas the mean CUSUM detects the same shift
application of our proposed robust CUSUM control at sample 15. The other three charts—that is, median,
charting structures using a real data set about patients’ HL, and TM CUSUMs—were able to absorb the
waiting times (in minutes) for a colonoscopy pro- variance contamination without giving a false alarm.
cedure in a regional medical center (cf. Jones-Farmer In case of location contamination, we can see that
et al. 2009). The data, consisting of 30 samples of size the plotting statistic (Sþ) for all of the charts is inflated
5, are presented in Figure 9a. All of the proposed a bit, but no out-of-control signal is received for any
charts are applied on these data and the outputs of chart (cf. Figure 11). Figure 12 is an interesting one

FIGURE 10 For 10% variance-contaminated data: (a) patients’ FIGURE 11 For 5% location-contaminated data: (a) patients’
waiting times; (b) output of mean CUSUM; (c) output of median waiting times; (b) output of mean CUSUM; (c) output of median
CUSUM; (d) output of mid-range CUSUM; (e) output of HL CUSUM; CUSUM; (d) output of mid-range CUSUM; (e) output of HL CUSUM;
and (f) output of TM CUSUM. and (f) output of TM CUSUM.

H. Z. Nazir et al. 222

 occasionally be present. The charts, which have a
robust design structure, are used to cope with such
environments. This study presents different robust
design structures for CUSUM-type charts and evaluates
their performance in different environments. The find-
ings of the article are that the mean CUSUM control
chart performs efficiently in uncontaminated environ-
ments and the TM and HL CUSUM charts are good com-
pared to the mean CUSUM chart in this situation. The
median CUSUM chart outperforms the other charts in
the presence of special cause environment and outliers.
The TM CUSUM chart is an alternative to the median
chart and is highly efficient in the presence of outliers.
It is also a good option in non-normal parent environ-
ments. Finally, the study concludes that the TM CUSUM
chart is the best choice for controlling the location para-
meter of a process in normal, non-normal, special
cause, and outlier environments.

ACKNOWLEDGMENT
The author Muhammad Riaz is indebted to King
Fahd University of Petroleum and Minerals, Dhahran,
FIGURE 12 For special cause environment with u ¼ 0.05: (a) Saudi Arabia, for providing excellent research
patients’ waiting times; (b) output of mean CUSUM; (c) output
of median CUSUM; (d) output of mid-range CUSUM; (e) output facilities through project SB111008.
of HL CUSUM; and (f) output of TM CUSUM.

ABOUT THE AUTHORS

where we can see that there are three outliers
present in the data set (cf. Figure 12a) at samples 4, Hafiz Zafar Nazir obtained his M.Sc. in statistics
27, and 28. Median, HL, and TM CUSUMs assimilated from the Department of Statistics, Quaid-i-Azam
these outliers, but the mean and mid-range CUSUM University, Islamabad, Pakistan, in 2006, and M.Phil
charts both gave the out-of-control signals at sample in statistics from the Department of Statistics,
27. The findings of this example are exactly on the Quaid-i-Azam University, Islamabad, Pakistan, in
same pattern as indicated by the ARL values for the 2008. Currently he is pursuing his Ph.D. in statistics
proposed charts in Tables 1–13. from the Institute for Business and Industrial
Statistics, University of Amsterdam, The Netherlands.
He has served as a lecturer in the Department of
SUMMARY AND CONCLUSIONS Statistics, University of Sargodha, Sargodha, Pakistan,
Common cause variations are an inherent part of from September 2009 to the present. His current
any process, and with timely monitoring, evaluation, research interests include statistical process control
and identification of sources of special cause varia- and nonparametric techniques. His e-mail address
tions, the quality of the output of the process can is [email protected].
be improved and waste (of time and cost) can be Muhammad Riaz obtained Ph.D. in statistics from
reduced significantly. Control charts are widely used the Institute of Business and Industrial Statistics,
to monitor a process. To monitor the location and University of Amsterdam, The Netherlands, in 2008.
dispersion parameters of the process, two main types He holds the position of assistant professor in the
of charts, Shewhart-type control charts and memory Department of Statistics, Quaid-i-Azam University,
control charts (CUSUM and EWMA), are used. In Islamabad, Pakistan, and the Department of
practice, special cause variations and outliers may Mathematics and Statistics, King Fahad University of

223 CUSUM Control Charting

Petroleum and Minerals, Dhahran, Saudi Arabia. His Langenberg, P., Iglewicz, B. (1986). Trimmed mean and R charts. Journal
of Quality Technology, 18(3):152–161.
current research interests include statistical process Lee, M. H. (2011). Economic design of cumulative sum control chart for
control, nonparametric techniques, and experimental non-normally correlated data. Matematika, 27(1):79–96.
Lehmann, E. L. (1983). Theory of Point Estimation New York: Wiley.
design.
Li, S. Y., Tang, L. C., Ng, S. H. (2010). Non-parametric CUSUM and
Ronald J. M. M. Does is professor of industrial EWMA control charts for detecting mean shifts. Journal of Quality
statistics at the University of Amsterdam. He is also Technology, 42(2):209–226.
Lucas, J. M., Crosier, R. B. (1982). Fast initial response for CUSUM
managing director of the Institute for Business and quality-control scheme. Technometrics, 24:199–205.
Industrial Statistics, which operates as an inde- MacEachern, S. N., Rao, Y., Wu, C. (2007). A robust-likelihood cumulat-
ive sum chart. Journal of the American Statistical Association,
pendent consultancy firm within the University of
102(480):1440–1447.
Amsterdam. Furthermore, he is director of the McDonald, D. (1990). A CUSUM procedure based on sequential ranks.
Graduate School of Executive Programmes at the Naval Research Logistics, 37:627–646.
Midi, H., Shabbak, A. (2011). Robust multivariate control charts to detect
Amsterdam Business School. His current research small shifts in mean. Mathematical Problems in Engineering, 2011,
activities lie in the design of control charts for DOI:10.1155=2011=923463.
Montgomery, D. C. (2009). Introduction to Statistical Quality Control,
nonstandard situations, health care engineering,
6th ed. New York: Wiley.
and Lean Six Sigma methods. Moustafa, A. (2009). A control chart based on robust estimators for
Nasir Abbas is serving as Assistant Professor at monitoring the process mean of a quality characteristic. International
Journal of Quality and Reliability Management, 26(5):480–496.
Department of Statistics, University of Sargodha Moustafa, A., Mokhtar, B. A. (1999). New robust statistical process
Pakistan. He got his M.Sc. in Statistics from the control chart for location. Quality Engineering, 12(2):149–159.
Page, E. S. (1954). Continuous inspection schemes. Biometrika, 41:100–115.
Department of Statistics Quaid-i-Azam University
Reynolds, M. R., Jr., Stoumbos, Z. G. (2010). Robust CUSUM charts for
Islamabad Pakistan in 2009; M.Phil. in Statistics from monitoring the process mean and variance. Quality and Reliability
the Department of statistics Quaid-i-Azam University Engineering International, 26:453–473.
Riaz, M. (2008). A dispersion control chart. Communications in Statistics -
Islamabad Pakistan in 2011 and Ph.D. in Industrial Simulation and Computation, 37:1239–1261.
Statistics from the Institute for Business and Roberts, S. W. (1959). Control chart tests based on geometric moving
averages. Technometrics, 1:239–250.
Industrial Statistics University of Amsterdam The
Rocke, D. M. (1992). X Q and RQ charts: Robust control charts.
Netherlands in 2012. He served as Assistant Census The Statistician, 41:97–104.
Commissioner in Pakistan Bureau of Statistics during Rousseeuw, P. J. (1991). Tutorial to robust statistics. Journal of
Chemometrics, 5:1–20.
July 2011 – January 2013. His current research inter- Schoonhoven, M., Nazir, H. Z., Riaz, M., Does, R. J. M. M. (2011a).
ests include Quality Control particularly control Robust location estimators for the X control chart. Journal of Quality
Technology, 43(4):363–379.
charting methodologies under parametric and non-
Schoonhoven, M., Riaz, M., Does, R. J. M. M. (2011b). Design and
parametric environments. analysis of control charts for standard deviation with estimated
parameters. Journal of Quality Technology, 43(4):307–333.
Tatum, L. G. (1997). Robust estimation of the process standard deviation
for control charts. Technometrics, 39:127–141.
Tukey, J. W. (1977). Exploratory Data Analysis Reading, MA: Addison
REFERENCES Wesley.
Vining, G. (2009). Technical advice: Phase I and phase II control charts.
Amin, R. W., Searcy, A. J. (1991). A nonparametric exponentially Quality Engineering, 21:478–479.
weighted moving average control scheme. Communications in Wang, T., Li, Y., Cui, H. (2007). On weighted randomly trimmed means.
Statistics - Simulation and Computation, 20(4):1049–1072. Journal of Systems Science and Complexity, 20:47–65.
Dixon, W. J., Massey, F. J. (1969). Introduction to Statistical Analysis, 3rd Weisberg, H. F. (1992). Central Tendency and Variability Sage University
ed. New York: McGraw-Hill. Paper Series on Quantitative Application in Social Sciences, Series
Ferrell, E. B. (1953). Control charts using midranges and medians. no. 07–038. A. Virding, Ed. Newbury Park, CA: Sage.
Industrial Quality Control, 9:30–34. Woodall, W. H., Montgomery, D. C. (1999). Research issues and ideas
Graham, M. A., Mukherjee, A., Chakraborti, S. (2012). Distribution-free in statistical process control. Journal of Quality Technology, 31:
exponentially weighted moving average control charts for monitoring 376–386.
unknown location. Computational Statistics and Data Analysis, Wu, C., Zhao, Y., Wang, Z. (2002). The median absolute deviation and
56:2539–2561. their applications to Shewhart Xbar control charts. Communications
Hawkins, D. (1993). Robustification of cumulative sum charts by in Statistics - Simulation and Computation, 31(3):425–442.
Winsorization. Journal of Quality Technology, 25(4):248–261. Yang, L., Pai, S., Wang, Y. R. (2010). A novel CUSUM median control
Hettmansperger, T. P., McKean, J. W. (1998). Robust Nonparametric chart. Paper presented at the International Multiconference of
Statistical Methods New York: Wiley. Engineers and Computer Scientists, March 17–19, 2010, Hong Kong.
Janacek, G. J., Miekle, S. C. (1997). Control charts based on medians. The Yang, S. F., Cheng, S. W. (2011). A new non-parametric CUSUM mean
Statistician, 46(1):19–31. chart. Quality and Reliability Engineering International, 27:867–875.
Jones-Farmer, L. A., Jordan, V., Champ, C. W. (2009). Distribution-free Yang, S. F., Lin, J. S., Cheng, S. W. (2011). A new nonparametric
phase I charts for subgroup location. Journal of Quality Technology, EWMA sign control chart. Expert Systems with Applications,
41(3):304–316. 38:6239–6243.

H. Z. Nazir et al. 224

Copyright of Quality Engineering is the property of Taylor & Francis Ltd and its content may
not be copied or emailed to multiple sites or posted to a listserv without the copyright holder's
express written permission. However, users may print, download, or email articles for
individual use.