Professional Documents
Culture Documents
Computation
Faijun Nahar Mim, Michael B. C. Khoo, Sajal Saha & Khai Wah Khaw
To cite this article: Faijun Nahar Mim, Michael B. C. Khoo, Sajal Saha & Khai Wah Khaw
(2023) A side-sensitive group runs median control chart with measurement errors,
Communications in Statistics - Simulation and Computation, 52:4, 1660-1678, DOI:
10.1080/03610918.2021.1888997
1. Introduction
A control chart is an important tool in Statistical Quality Control to monitor process quality in man-
ufacturing and service industries. The role of a control chart is to detect process shifts as early as
possible so that corrective actions can be taken to promptly bring back the out-of-control process to
an in-control situation again. Since the Shewhart control chart was introduced by W.A. Shewhart in
the 1920s, enormous research on control charts has resulted in different types of control charts in
the literature, among others are the synthetic and group runs (GR) charts. The synthetic X chart was
suggested by Wu and Spedding (2000) and a literature review of these types of charts was recently
provided by Rakitzis et al. (2019). Owing to the effectiveness of the synthetic X chart, numerous
extensions have been proposed, for example, see Zhang et al. (2011), Khoo et al. (2010), Haq and
Khoo (2016), Lee and Khoo (2017), Wan et al. (2018), Khaw et al. (2019), Haq (2019), Tran,
Tran, and Rakitzis (2019), Shongwe, Malela-Majika, and Castagliola (2020), and Shongwe and
Malela-Majika (2020), to name a few.
In addition to the synthetic chart, the GR X chart was proposed by Gadre and Rattihalli
(2004), where they integrated the Shewhart X and conforming run length (CRL) sub-charts to
detect mean shifts in the process. The GR X chart signals an out-of-control if CRL1 < L or both
CONTACT Michael B. C. Khoo mkbc@usm.my School of Mathematical Sciences, Universiti Sains Malaysia, Minden,
Penang, Malaysia.
ß 2021 Taylor & Francis Group, LLC
COMMUNICATIONS IN STATISTICS - SIMULATION AND COMPUTATIONV
R
1661
CRLr and CRLrþ1 (for r > 1) are less than L, where L is the lower limit of the CRL sub-chart (of
the GR X chart). Here, CRLr represents the number of conforming samples, inspected between
the (r 1)th and rth non-conforming samples, including the rth non-conforming sample. In order
to improve the performance of the GR X chart, Gadre and Rattihalli (2007) added the side-sensi-
tivity feature into the GR X chart and introduced a new chart, called the side-sensitive group
runs (SSGR) X chart. The side-sensitivity feature works based on the fact that two points that fall
beyond the control limits of a chart should be on the same side of the center line, in order for
the chart to issue a signal. That is, in the case of the SSGR X chart, an out-of-control is signaled
by the chart when the sample means X that correspond to the two successive CRLs that contrib-
ute to the out-of-control signal fall on the same side of the center line of the Shewhart X sub-
chart. The SSGR X chart was shown to surpass the existing Shewhart X, synthetic X and GR X
charts. The combined SSGR and double sampling (DS) charts, called the SSGRDS chart was pro-
posed by Khoo et al. (2015).
On similar lines, by integrating the side sensitivity feature with the modified group runs
(MGR) chart proposed by Gadre and Rattihalli (2006), Gadre, Joshi, and Rattihalli (2010) devel-
oped the SSMGR chart, which outperforms all existing GR charts. The SSMGR chart tiggers an
out-of-control signal when CRL1 < L or for r > 1, CRLr L1 and CRLrþ1 L, where L1 and L
are the lower limits of the CRL sub-chart. Following the advantage of the SSMGR chart, Saha
et al. (2018) suggested the side-sensitive modified group runs double sampling (SSMGRDS) chart
for detecting process mean shifts. You et al. (2016) presented the SSGR chart for monitoring the
process coefficient of variation. Mim et al. (2019) developed the side-sensitive modified group
runs (SSMGR) chart with auxiliary information for the process mean. More recently, Motsepa
et al. (2020) studied the effect of parameter estimation on the side-sensitive double sampling X
chart in detecting changes in the process mean.
The main drawback of mean type charts is that they are sensitive to outliers. In practice, it is
common for outliers to be present in industrial processes. If the X chart is used to monitor the
process mean in such situations, it will produce a high false alarm rate. Similar to the sample
mean ðX Þ, the sample median is also a measure of central tendency. However, the sample
median is a more robust estimator of central tendency than the sample mean, which makes the
former more desirable in situations when outliers are present. Therefore, the sample median is
often used in place of the sample mean to monitor the process mean when outliers are prone to
occur. All existing papers on median type charts (see for example, Castagliola, Maravelakis, and
Figueiredo (2016), Hu et al. (2019), Tran (2017) and Saha et al. (2019), to name a few) consid-
ered the monitoring of the process mean, instead of the median. This is because in the monitor-
ing of central tendency, control charts are usually designed to monitor the process mean. Median
type charts are merely used as substitutes of mean type charts to offer robustness to outliers that
are present in a process. The use of median type charts over mean type charts to monitor the
process mean offers advantages when dealing with slightly to moderately skewed populations,
besides providing ease of computation as the computation of the sample median does not require
all individual observations to be accurately measured (unlike computing the sample mean).
Another advantage of the median chart is that it maintains the chart’s sensitivity in the presence
of outliers because the limits computed from the sample medians are less stretched by outlying
values compared to that computed from the sample means (in the case of the X chart).
Khoo (2005) presented a median chart to monitor shifts in the process mean. Park (2009)
introduced a new median chart, where the control limits are computed using the bootstrap tech-
nique. Castagliola and Figueiredo (2013) proposed a median chart when the process mean and
standard deviation are estimated. Costa and Machado (2013) integrated the side-sensitive (SS)
synthetic feature into the median chart to propose the SS synthetic X ~ chart. The SS synthetic X ~
chart was shown to surpass the existing X ~ chart, in terms of the ARL criterion. Castagliola,
Maravelakis, and Figueiredo (2016) proposed the EWMA median chart with estimated
1662 F. N. MIM ET AL.
parameters. The synthetic X ~ chart was presented in Tran, Tran, and Rakitzis (2019), where the
chart’s statistical properties were evaluated using the Markov chain method. Recently, the non-
parametric corrected sample based SS synthetic and nonparametric sample based SSGR charts to
detect shifts in the process median were suggested by Gadre and Kakade (2019).
Measures taken on inspected units are often found to be contaminated in many industrial
applications by the inspector or the measuring equipment leading to errors in the measurements.
Most of the research works on control charts in the literature assume that measurement error
does not exist in the values of inspected units. However, this assumption is often untrue in prac-
tice as measurement errors are difficult to prevent. A literature review on the effect of measure-
ment errors on the performance of control charts in numerous process monitoring applications
was given in Maleki, Amiri, and Castagliola (2017). Measurement errors are usually encountered
in process monitoring and are distortions that hamper the efficiency of a control chart, leading to
erroneous decisions regarding the run length properties and performance of a control chart.
Bennett (1954) studied the effect of measurement errors on the Shewhart X chart, where he
showed that measurement errors affect the chart’s performance. Over the years, many researchers
have investigated the effects of measurement errors on the performances of different types of con-
trol charts, for instance, see Linna and Woodall (2001), Maravelakis, Panaretos, and Psarakis
(2004), Costa and Castagliola (2011), Hu et al. (2015), Tran, Castagliola, and Celano (2016),
Tran, Castagliola, and Balakrishnan (2017), Maleki, Amiri, and Castagliola (2017), Cheng and
Wang (2018), Saha et al. (2019), Lee et al. (2019) and Saha et al. (2021). Tran, Tran, and Rakitzis
(2019) presented a synthetic X ~ chart with measurement errors and they showed that the synthetic
~
X chart’s performance is severely affected by measurement errors. Riaz et al. (2019) presented
auxiliary information based mixed EWMA-cumulative sum (CUSUM) chart for detecting mean
shifts when measurement errors are present.
In this study, the SSGR X ~ chart for detecting shifts in the process mean is proposed.
Additionally, by adopting the measurement error model of Linna and Woodall (2001), the per-
formance of the SSGR X ~ chart in the presence of measurement errors is also studied. The rest of
this study is organized as follows: Section 2 discusses the design of the SSGR X~ chart and its stat-
istical performances using the Markov chain method. In Sec. 3, the performance of the SSGR X ~
chart is compared with that of existing charts when measurement error is not present. Section 4
presents the linear covariate error model for the SSGR X ~ chart. Section 5 discusses the SSGR X ~
chart and its properties in the presence of measurement errors. Section 6 investigates the effects
of measurement errors on the performance of the SSGR X ~ chart. An implementation of the
SSGR X ~ chart with measurement errors is illustrated in Sec. 7. Finally, conclusions and recom-
mendations are drawn in Sec. 8.
~ chart
2. Design of the SSGR X
Assume that fXi, 1 , Xi, 2 , :::, Xi, n g are the Phase-II samples, each of size n, where i (¼ 1, 2, … ) denotes
the sample number. We assume that there is independence within and between samples and Xij fol-
lows a normal distribution, i.e., Xij Nðl0 þ dr, r2 Þ, for j ¼ 1, 2, … , n. Here, l0 and r are the in-
control mean and standard deviation, respectively, while d ¼ jl1 l r
0j
quantifies the standardized
mean shift from l0 to the out-of-control mean, l1 : The process is in-control when d ¼ 0, otherwise,
it is out-of-control. Let X ~ i be the sample median of the ith sample of size n. Then
8
< Xi, ððnþ1Þ=2Þ , if n is odd
~
X i ¼ Xi, ðn=2Þ þ Xi, ðn=2þ1Þ , (1)
: , if n is even
2
where Xi, ð1Þ , Xi, ð2Þ , :::, Xi, ðnÞ is the ith ordered sample.
COMMUNICATIONS IN STATISTICS - SIMULATION AND COMPUTATIONV
R
1663
The cumulative distribution function (cdf) of X~ i when n is odd is given by (Castagliola and
Figueiredo 2013)
x l0 nþ1 nþ1
FX~ ðxÞ ¼ Fb U
i
d , , (2)
r0 2 2
Step 1. Take a sample of size n and compute the sample median, X ~ i , for i ¼ 1, 2, … , using
Equation (1).
~ i 2 [LCL, UCL], the ith sample is conforming and the process is considered as in-con-
Step 2. If X
trol, then return to Step 1 to take the next sample. Otherwise, sample i is non-conforming and
the control flow proceeds to Step 3.
Step 3. If Y1 L or for r > 1, Yr L and Yrþ1 L, and that both Yr and Yrþ1 fall on the same
side of the X ~ sub-chart, the process is declared as out-of-control and proceed to Step 4.
Otherwise, return to Step 1.
Step 4. Investigate the underlying process and identify the presence of assignable causes. Take
corrective actions to bring back the out-of-control process into the in-control situation again.
~ sub-chart. Then
Let h be the probability that a sample is declared as conforming on the X
h ¼ Pr X ~ i 2 ½LCL, UCL
1664 F. N. MIM ET AL.
It follows from the above discussion that the probability of declaring a sample as non-con-
forming on the X ~ sub-chart is computed as
h0 ¼ Pr X ~ i 62 ½LCL, UCL
¼ 1 h, (7)
where h is obtained from Equation (6). Let
a ¼ Pr X ~ > UCLjX~ 62 ½LCL, UCL
1 Pr X ~ UCL
¼
h'
1 Fb ðUðK dÞj nþ1
2 , 2 Þ
nþ1
¼ 0 : (8)
h
A similar Markov chain approach to that of the SSGR X chart of Gadre and Rattihalli (2007)
is used to derive the average run length (ARL) and standard deviation of the run length (SDRL)
formulae for the SSGR X ~ chart. The different possible outcomes for a sampling stage in the SSGR
~
X chart with their respective notations are described as follows:
(a) 1 is an imaginary case representing the head start which indicates a non-conforming sam-
ple at time zero ðY0 LÞ and the process mean shift is either an increasing shift or a
decreasing shift.
(b) 0 represents a conforming sample with Yr > L.
(c) 1 represents a non-conforming sample with Yr L and the process mean shift is either an
increasing shift or a decreasing shift.
(d) 1 represents a non-conforming sample with Yr L and the process mean shift is a decreas-
ing shift.
(e) 1 represents a non-conforming sample with Yr L and the process mean shift is an
increasing shift.
For illustrating the construction of the transition probability matrix (tpm) for the Markov
chain model, L ¼ 2 is considered. Consequently, the SSGR X ~ chart signals an out-of-control if
Y1 2 or Yr 2 and Yrþ1 2, for r ¼ 2, 3, … , and that both Yr and Yrþ1 are having shifts on
the same side of the X~ sub-chart. Table 1 displays the Markov chain states for L ¼ 2, where states
1 to 9 are transient states. The transient states 1 9 can be explained as follows (Gadre and
Rattihalli 2007):
COMMUNICATIONS IN STATISTICS - SIMULATION AND COMPUTATIONV
R
1665
Table 1. Markov chain states for the SSGR X chart with L ¼ 2.
State State representation State State representation State State representation
1 1 5 0010 9 10
2 10 6 1 10 Signal
3 00 7 10
4 001 8 1
1. A sequence starting with 1 and followed by at most (L 1) zeros. There are L states of this
type (states 1 and 2 in Table 1).
2. A sequence of L zeros. There is one such state (state 3 in Table 1).
3. A sequence of L zeros followed by 1 and is further added by at most (L – 1) zeros. There are
L states of this type (states 4 and 5 in Table 1).
4. A sequence of 1 or 1 followed by at most (L 1) zeros. There are 2L states of this type
(states 6 to 9 in Table 1).
where h, h0 and a are defined in Equations (6), (7) and (8), respectively.
The ARL and SDRL values for the SSGR X ~ chart are computed as follows:
ARL ¼ U1 (9)
and
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
SDRL ¼ U2 U12 þ U1 , (10)
where
Uv ¼ v! qT ðI QÞ–v Qv1 1, for v ¼ 1, 2: (11)
T
Note that I is a (4 L þ 1) (4 L þ 1) identity matrix and q ¼ ð1, 0, :::, 0Þ is a (4 L þ 1) 1 initial
probability vector.
The optimal parameters K0 and L0 of the SSGR X ~ chart in minimizing the out-of-control ARL
(ARL1) value for the shift size d are obtained as follows:
ðK 0 , L0 Þ ¼ argmin ARL1 ðn, K, L, dÞ (12a)
ðK , LÞ
subject to
ARLðn, K, L, d ¼ 0Þ ¼ ARL0 : (12b)
Here, ARL0 denotes the in-control ARL. The algorithm in computing the optimal parameters
~ chart in minimizing the ARL1 value for the shift size d based on the
(K0 , L0 ) of the SSGR X
objective function in Equation (12a) is as follows:
1666 F. N. MIM ET AL.
Table 2. The tpm of the SSGR X chart with L ¼ 2.
Next State (t)
Current State(s) 1 2 3 4 5 6 7 8 9 10
1 0 h 0 h’ 0 0 0 0 0 h’
2 0 0 h 0 0 0 0 0 0 h’
3 0 0 h 0 0 0 0 0 0 0
4 0 0 0 0 h a0 h 0 ð1 a0 Þh 0 0
5 0 0 h 0 0 a0 h 0 ð1 a0 Þh 0 0
6 0 0 0 0 0 0 h ð1 a0 Þh 0 a0 h
7 0 0 h 0 0 0 0 ð1 a0 Þh 0 a0 h
8 0 0 0 0 0 a0 h 0 0 h ð1 a0 Þh
9 0 0 h 0 0 a0 h 0 0 0 ð1 a0 Þh
10 0 0 0 0 0 0 0 0 0 1
~ chart
3. Performance analyses of the SSGR X
An optimization program in Matlab incorporating Steps 1 6 in Sec. 2 is written to compute the
optimal parameters (K0 , L0 ) that minimize the ARL1 values of the SSGR X ~ chart for shift sizes d
2 {0.1, 0.2, 0.3, 0.5, 0.7, 1, 1.5, 2}, based on ARL0 ¼ 370.4 and n 2 {3, 5, 7}. The chart’s corre-
sponding out-of-control SDRL (SDRL1) values are also computed for these shift sizes. Table 3
presents the optimal parameters (K0 , L0 ) of the SSGR X ~ chart computed based on the above input
parameters. For example, when ARL0 ¼ 370.4 and n ¼ 5, the optimal parameters that minimize
ARL1 for the shift d ¼ 0.3 are (K0 , L0 ) ¼ (1.2309, 26) (see Table 3). These optimal parameters give
ARL1 and SDRL1 values of 51.5 and 81, respectively (see Table 4).
The performance of the SSGR X ~ chart is compared with that of existing median type charts,
such as the synthetic X,~ SS synthetic X ~ and EWMA X ~ charts, in terms of the ARL and SDRL cri-
teria. Tran, Tran, and Rakitzis (2019) showed that the synthetic X ~ chart outperforms the
Shewhart X ~ chart. Therefore, the Shewhart X ~ chart is not considered in the comparison. Table 4
shows the ARL1 and SDRL1 values for the SSGR X, ~ synthetic X,~ SS synthetic X~ and EWMA X ~
charts. From Table 4, it is seen that the SSGR X ~ chart outperforms the synthetic X ~ chart, for all
shift sizes, d and sample sizes, n, as the former has smaller ARL1 values than the latter for the
same (d, n) combination. For instance, when n 2 {3, 5, 7} and d ¼ 0.5, ARL1 2 {24.9, 13.3, 8.5}
for the SSGR X ~ chart, while ARL1 2 {46, 25.3, 16} for the synthetic X ~ chart (see Table 4), where
the ARL1 values of the former are lower than that of the latter. Furthermore, it is noticed in
Table 4 that the SSGR X ~ chart beats the SS synthetic X ~ chart, in terms of the ARL1 criterion
when d 0.3, even though both charts have equal ARL1 performances when d ¼ 2: For example,
when n ¼ 5 and d ¼ 0.3, ARL1 ¼ 51.5 for the SSGR X ~ chart is lower than ARL1 ¼ 56.8 for the SS
synthetic X ~ chart. In comparison to the EWMA X ~ chart, the SSGR X ~ chart outperforms the
EWMA X ~ chart for moderate and large shift sizes, d. The EWMA X ~ chart prevails in detecting
small shifts. For example, when n ¼ 5 and the shift sizes are small, i.e., d 2 {0.1, 0.2, 0.3, 0.5},
ARL1 2 {206.9, 81.9, 37.3, 12.6} for the EWMA X ~ chart, which are less than that for the SSGR X ~
chart, given by ARL1 2 {253.4, 116, 51.5, 13.3} (see Table 4). Therefore, for small shifts, the
COMMUNICATIONS IN STATISTICS - SIMULATION AND COMPUTATIONV
R
1667
Table 3. Optimal parameters (K’, L’ ) of the SSGR X chart in minimizing ARL1 for the shift size d when ARL0 ¼ 370.4 and n 2
{3, 5 7}.
n¼3 n¼5 n¼7
d K’ L’ K’ L’ K’ L’
0.1 1.6694 60 1.3239 55 1.1351 56
0.2 1.6296 46 1.2821 39 1.0861 35
0.3 1.5781 33 1.2309 26 1.0353 22
0.5 1.4796 18 1.1384 13 0.9434 10
0.7 1.3951 11 1.0503 7 0.8799 6
1 1.2853 6 0.9657 4 0.7882 3
1.5 1.1511 3 0.8538 2 0.7314 2
2 1.0679 2 0.8538 2 0.7314 2
Table 4. ARL1 and SDRL1 values for the SSGR X , synthetic X , side-sensitive (SS) synthetic X and EWMA X charts when ARL0
¼ 370.4 and n 2 {3, 5, 7}.
SSGR X Synthetic X SS synthetic X EWMA X
d ARL1 SDRL1 ARL1 SDRL1 ARL1 SDRL1 ARL1 SDRL1
n¼3
0.1 287.6 463.5 318.9 413.0 300.78 297.30 247.6 244.3
0.2 161.6 258.3 217.2 282.5 193.13 187.61 117.9 114.2
0.3 83.6 132.5 131.5 171.7 117.87 111.18 58.6 54.9
0.5 24.9 38.5 46.0 60.1 46.85 40.52 20.3 16.7
0.7 9.6 14.0 18.3 23.4 22.84 17.75 10.1 7.0
1 3.6 4.4 6.3 7.4 10.41 6.77 5.3 2.8
1.5 1.5 1.2 2.1 1.8 4.40 2.23 2.8 1.3
2 1.1 0.4 1.3 0.7 2.72 0.98 1.8 0.9
n¼5
0.1 253.4 406.7 294.6 382.1 275.86 271.31 206.9 203.5
0.2 116.0 184.5 170.0 221.8 148.05 144.87 81.9 78.1
0.3 51.5 81.0 87.8 114.8 78.86 74.75 37.3 33.6
0.5 13.3 19.9 25.3 32.6 27.00 24.69 12.6 9.4
0.7 5.1 6.9 9.4 11.4 11.91 9.89 6.6 3.9
1 2.1 2.1 3.3 3.5 5.04 3.50 3.7 1.7
1.5 1.2 0.6 1.4 0.8 2.51 0.91 2.0 1.0
2 1.0 0.1 1.1 0.3 2.08 0.29 1.3 0.6
n¼7
0.1 224.9 362.1 272.5 354.0 248.69 244.68 176.7 173.2
0.2 87.5 138.9 136.2 177.8 117.68 114.51 61.5 57.8
0.3 35.1 54.7 62.7 81.9 57.79 54.35 26.8 23.2
0.5 8.5 12.2 16.0 20.3 18.37 16.50 9.3 6.2
0.7 3.4 4.1 5.9 6.8 7.99 6.34 5.1 2.7
1 1.6 1.3 2.2 2.0 3.70 2.03 3.0 1.4
1.5 1.1 0.3 1.2 0.5 2.20 0.49 1.6 0.8
2 1.0 0.1 1.0 0.1 2.01 0.11 1.1 0.3
EWMA X ~ chart is quicker than the SSGR X ~ chart in detecting the shifts, while the opposite is
true for moderate and large shifts.
In terms of the SDRL criterion, it is noticeable from Table 4 that the SSGR X ~ chart prevails
~
over the synthetic X chart for all shift sizes, d, except d ¼ 0.1, as the former has smaller SDRL1
values than the latter for d 2 {0.2, 0.3, 0.5, 0.7, 1, 1.5, 2}. A smaller SDRL value translates into a
smaller spread in the run length distribution of the chart, which is a desirable property. As an
example, when n ¼ 7 and d 2 {0.2, 0.3, 0.5, 0.7, 1, 1.5, 2}, SDRL1 2 {138.9, 54.7, 12.2, 4.1, 1.3,
0.3, 0.1} for the SSGR X ~ chart, while that for the synthetic X ~ chart are SDRL1 2 {177.8, 81.9,
20.3, 6.8, 2, 0.5, 0.1}, where the SDRL1 values of the former are smaller than those of the latter.
The SS synthetic X ~ chart loses out to the SSGR X ~ chart, in terms of the SDRL1 creterion for
d 0.5, but for d ¼ 2, both charts have equal performances (see Table 4). For example, when
n ¼ 3, d ¼ 1, SDRL1 ¼ 5.5 for the SS synthetic X ~ chart is larger than SDRL1 ¼ 4.4 for the SSGR
~ chart.
X
1668 F. N. MIM ET AL.
Table 5. Optimal parameters (K’, L’ ) of the SSGR X chart in minimizing ARL1 for the shift size d when ARL0 ¼ 370.4 and n 2
{4, 6, 8}.
n¼4 n¼6 n¼8
d K’ L’ K’ L’ K’ L’
0.1 1.2913 35 1.0679 28 0.9151 20
0.2 1.2799 32 1.0679 28 0.9151 20
0.3 1.2629 28 1.0545 24 0.9100 19
0.5 1.1702 14 0.9747 12 0.8186 8
0.7 1.0483 6 0.9262 8 0.7652 5
1 0.9846 4 0.7971 3 0.7049 3
1.5 0.8710 2 0.7394 2 0.5613 1
2 0.8710 2 0.7394 2 0.5613 1
The EWMA X ~ chart performs better than the SSGR X ~ chart, for small and moderate mean
shifts, but for large shifts, the latter outperforms the former. For example, when n ¼ 5 and the
shifts are large, say d 2 {1.5, 2}, SDRL1 2 {0.6, 0.1} for the SSGR X ~ chart, which are less than
that for the EWMA X ~ chart, given by SDRL1 2 {1, 0.6} (see Table 4). However, for the same
sample size when the shifts are small and moderate, say d 2 {0.1, 0.2, 0.3, 0.5, 0.7, 1}, SDRL1 2
{406.7, 184.5, 81, 19.9, 6.9, 2.1} for the SSGR X ~ chart but that for the EWMA X ~ chart are SDRL1
2 {203.5, 78.1, 33.6, 9.4, 3.9, 1.7}, where the latter has smaller SDRL1 values than the former.
A program in Matlab incorporating Steps 1 6 in Sec. 2 is written to compute the optimal
parameters (K0 , L0 ) that minimize the ARL1 values of the SSGR X ~ chart for shift sizes d 2 {0.1,
0.2, 0.3, 0.5, 0.7, 1, 1.5, 2}, based on ARL0 ¼ 370.4 and n ¼ {3, 5, 7}.
When the sample size (n) is even, a simulation using the SAS program is conducted, based on
ARL0 ¼ 370.4, to investigate the performance of the SSGR X ~ chart. The even sample sizes, n 2
0 0
{4, 6, 8} are considered. The optimal parameters (K , L ) of the SSGR X ~ chart are presented in
Table 5, while the corresponding ARL1 and SDRL1 values of the SSGR X ~ and competing charts
are presented in Table 6. A similar trend to the case of odd sample sizes is observed when the
SSGR X ~ chart is compared with the synthetic X, ~ SS synthetic X ~ and EWMA X ~ charts, in terms
of the ARL and SDRL criteria.
In Sec. 4 and 5, the linear covariate error model for the X ~ chart and the SSGR X ~ chart with
measurement errors will be discussed for the case of odd sample sizes. For the case of even sam-
ple sizes, a similar approach can be employed to derive the formulae for conducting the same
type of analyses.
~ chart
4. Linear covariate error model for X
Most of the models on measurement errors in the literature are the additive, multiplicative, two
component and four component ones. Among these models, the simple yet intitutive additive
model is the most popular (Maleki, Amiri, and Castagliola 2017). For this reason, this study
adopts the additive error model. Futhermore, the “multiple measurements” approach proposed by
Linna and Woodall (2001) is one of the most popular remedial approaches to compensate for the
negative effects of measurement errors. In this approach, multiple measurements (say m) per
item in each sample is taken, instead of taking a single measurement, and then the mean of the
multiple measured values for each item is computed. Consequently, the variation in the measure-
ment errors as a result of the multiple measurements approach will become smaller than that
based on the single measurement approach.
In this study, a linear covariate error model suggested by Linna and Woodall (2001) will be
adopted for the SSGR X ~ chart. It is assumed that at time i ¼ 1, 2, … , the quality characteristics
fXi, 1 , Xi, 2 , :::, Xi, n g, for n 1 are independent normal random variables, where Xi, j Nðl ~ 0þ
dr0 , r20 Þ: By considering a similar assumption to that in Linna and Woodall (2001), Xi, j is not
COMMUNICATIONS IN STATISTICS - SIMULATION AND COMPUTATIONV
R
1669
Table 6. ARL1 and SDRL1 values for the SSGR X , synthetic X , side-sensitive (SS) synthetic X and EWMA X charts when ARL0
¼ 370.4 and n 2 {4, 6, 8}.
SSGR X Synthetic X SS synthetic X EWMA X
d ARL1 SDRL1 ARL1 SDRL1 ARL1 SDRL1 ARL1 SDRL1
n¼4
0.1 253.81 393.87 296.53 365.70 268.72 259.87 97.05 63.73
0.2 119.73 186.26 174.05 220.22 146.69 142.17 40.64 24.27
0.3 52.86 83.43 89.80 117.63 79.43 74.98 23.16 13.30
0.5 13.89 20.84 26.61 34.32 27.28 24.60 10.92 5.80
0.7 15.39 7.54 9.89 12.23 12.37 10.04 6.93 2.75
1 2.20 2.29 3.51 3.92 5.22 3.51 3.84 1.82
1.5 1.19 0.6421 1.43 1.00 2.57 1.00 2.24 0.76
2 1.02 0.1543 1.06 0.29 2.09 0.31 1.64 0.55
n¼6
0.1 215.59 328.90 272.70 355.06 248.97 244.53 80.34 49.63
0.2 85.04 134.30 138.36 178.79 119.02 113.67 36.18 15.74
0.3 35.34 56.18 63.37 84.15 58.52 55.30 23.18 8.08
0.5 8.77 12.30 16.39 20.60 18.46 16.43 13.50 3.57
0.7 3.54 4.03 6.03 6.93 8.15 6.30 9.58 2.12
1 1.60 1.35 2.28 2.12 3.72 2.21 6.75 1.23
1.5 1.06 0.30 1.16 0.52 2.22 0.53 4.60 0.70
2 1.00 0.06 1.01 0.12 2.02 0.18 3.54 0.54
n¼8
0.1 205.03 307.95 250.89 321.82 227.92 222.09 68.90 45.68
0.2 70.01 109.40 112.75 147.19 97.34 93.65 28.97 13.40
0.3 25.83 40.04 47.12 62.18 46.00 43.64 18.10 6.65
0.5 6.17 8.52 11.53 14.44 13.73 12.03 10.35 2.83
0.7 2.58 2.86 2.21 1.63 6.04 4.34 7.31 1.66
1 1.35 0.88 1.76 1.42 3.00 1.46 5.15 0.96
1.5 1.03 0.33 1.06 0.28 2.09 0.32 3.52 0.58
2 1.00 0.05 1.00 0.06 2.00 0.07 2.80 0.44
directly measurable but it can only be computed from the outputs Xi, j, 1 , Xi, j, 2 , :::, Xi, j, m of a set
of m 1 measurements. Here, Xi, j, k (for k ¼ 1, 2, … , m) is defined by a linear covariate model
as
Xi, j, k ¼ A þ BXi, j þ ei, j, k , (13)
where A and B are two known constants and ei, j, k denotes the random error term having a nor-
mal distribution, i.e., ei, j, k ~ Nð0, r2M Þ: Note that ei, j, k occurs due to measurement inaccuracy and
it is independent of Xi, j : The smaller the value of the standard deviation of the measurement
error, rM , the higher is the measure of precision.
The mean X i, j of the observable quantities Xi, j, 1 , Xi, j, 2 , :::, Xi, j, m is computed as
Xm
i, j ¼ 1
X X
m k¼1 i, j, k
1X m
¼ ðA þ BXi, j þ ei, j, k Þ
m k¼1
1X m
¼ A þ BXi, j þ ei, j, k , (14)
m k¼1
and
rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
r2
r ¼ B2 r20 þ M : (16)
n m
The sample median X ~ i of the mean values X i, 1 , X
i, 2 , … ,X i, n g corresponding to sample i ¼ 1, 2,
… , is
8
>
< X if n is odd
i, ððnþ1Þ=2Þ
~
Xi ¼ X þX i, ðn=2þ1Þ , (17)
>
: i, ðn=2Þ if n is even
2
n
where X i, ð1Þ , X
i, ð2Þ , … , X
i, ðnÞ g is the ordered sample of the mean values for sample i (¼ 1, 2,
~ i is
… ). The cdf of X
xl n þ 1 n þ 2
FX~ ðxÞ ¼ Fb U ,
i r 2 2
! !
x A Bðl0 þ dr0 Þ n þ 1 n þ 2
¼ Fb U qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi2 ffi , : (18)
r 2 2
B2 r20 þ mM
0 0 1
dBr0 n þ 1 n þ 2
Fb @U@K qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi2 ffiA , : (20)
r 2 2
B2 r20 þ mM
COMMUNICATIONS IN STATISTICS - SIMULATION AND COMPUTATIONV
R
1671
Dividing the numerator and denominator in the argument of U () in Equation (20) by r0 , in
order to express h as a function of the precision error ratio, .. ¼ rrM0 , gives
h ¼ FX~ ðUCL Þ FX~ ðLCL Þ
0
i
0 i
1
dB n þ 1 n þ 2
¼ Fb @U@K qffiffiffiffiffiffiffiffiffiffiffiffiffiffiA ,
B2 þ g
2 2 2
m
0 0 1
dB n þ 1 n þ 2
Fb @U@K qffiffiffiffiffiffiffiffiffiffiffiffiffiffiA , : (21)
B2 þ gm
2 2 2
The ARL and SDRL values of the SSGR X ~ chart can be computed using Equations (9) and
(10), respectively. Note that the tpm Q is constructed in the same way as that of the SSGR X ~
chart elaborated in Sec. 2, except that h and a are replaced by h and a , respectively. Here, a is
computed as follows:
1 Pr X~ UCL
a ¼ : (22)
1h
The optimal parameters K and L of the SSGR X ~ chart in minimizing the ARL1 value for
the shift size d are obtained as follows:
ðK , L Þ ¼ argmin ARL1 ðn, K, L, g, B, dÞ (23a)
ð K , LÞ
subject to
ARL ðn, K, L, g, B, d ¼ 0Þ ¼ ARL0 (23b)
The optimization algorithm used in the optimal design of the SSGR X ~ chart in minimizing
the ARL1 value for the shift size d based on the objective function in Equation (23a) is given
as follows:
~ chart
6. Effects of measurement errors on the SSGR X
This section investigates the performance of the SSGR X~ chart in the presence of measurement
errors. Tables 7–9 give the optimal parameters ðK , L Þ of the SSGR X ~ chart in minimizing the
ARL1 value for the shift size, d, when measurement errors are present. Also recorded in these
tables are the minimum ARL1 values for the shift size d and their corresponding SDRL1 values
for the SSGR X ~ chart. The optimal parameters ðK , L Þ and the corresponding ARL1 values are
computed using the optimization procedure explained in Sec. 5 for different (m, n, B, g) combi-
nations satisfying ARL0 ¼ 370.4, where A ¼ 0 and d 2 {0.1, 0.2, 0.3, 0.5, 0.7, 1, 1.5, 2} are consid-
ered. For example, when the SSGR X ~ chart is designed to minimize ARL1 for d ¼ 0.3, the
1672 F. N. MIM ET AL.
Table 7. ðK , L , ARL1 , SDRL1 Þ values for the SSGR X chart for different values of g, d and n, where B ¼ 1 and m ¼ 1.
d g¼0 g ¼ 0:1 g ¼ 0:5 g ¼ 1:0
n¼3
0.1 1.6694, 60, 287.6, 463.5 1.6694, 60, 288.2, 464.6 1.6694, 60, 301.5, 485.5 1.6694, 60 324.5, 521.9
0.2 1.6296, 46, 161.6, 258.3 1.6296, 46, 162.7, 259.9 1.6422, 50, 184.9, 296.4 1.6538, 54, 231.7, 371.9
0.3 1.5781, 33, 83.6, 132.5 1.5828, 34, 84.4, 133.9 1.5960, 37, 103.0, 163.7 1.6228, 44, 149.4, 238.4
0.5 1.4796, 18, 24.9, 38.5 1.4796, 18, 25.3, 39.1 1.5052, 21, 33.5, 52.2 1.5520, 28, 59.1, 93.3
0.7 1.3951, 11, 9.6, 14.0 1.3951, 11, 9.8, 14.2 1.4242, 13, 13.3, 19.9 1.4796, 18, 25.6, 39.6
1 1.2853, 6, 3.6, 4.4 1.2853, 6, 3.6, 4.5 1.3138, 7, 4.8, 6.4 1.3783, 10, 9.4, 13.7
1.5 1.1511, 3, 1.5, 1.2 1.1511, 3, 1.5, 1.2 1.2079, 4, 1.8, 1.6 1.2509, 5, 3.1, 3.7
2 1.0679, 2, 1.1, 0.4 1.0679, 2, 1.1, 0.4 1.0679, 2, 1.2, 0.7 1.1511, 3, 1.7, 1.5
n¼5
0.1 1.3239, 55, 253.4, 406.7 1.3239, 55, 254.2, 408.0 1.3239, 55, 271.4, 435.0 1.3239, 55, 302.6, 484.0
0.2 1.2821, 39, 116.0, 184.5 1.2821, 39, 116.9, 186.1 1.2941, 43, 138.2, 220.4 1.3100, 49, 187.0, 299.2
0.3 1.2309, 26, 51.5, 81.0 1.2358, 27, 52.1, 82.0 1.2492, 30, 66.1, 104.5 1.2756, 37, 104.9, 166.6
0.5 1.1384, 13, 13.3, 19.9 1.1384, 13, 13.5, 20.2 1.1580, 15, 18.3, 27.8 1.2031, 21, 34.4, 53.7
0.7 1.0503, 7, 5.1, 6.9, 1.0503, 7, 5.2, 7.0 1.0867, 9, 7.0, 9.8 1.1384, 13, 13.7, 20.5
1 0.9657, 4, 2.1, 2.1, 0.9657, 4, 2.1, 2.2 1.0000, 5, 2.7, 3.0 1.0503, 7, 5.0, 6.6
1.5 0.8538, 2, 1.2, 0.6 0.8538, 2, 1.2, 0.6 0.9203, 3, 1.3, 0.8 0.9657, 4, 1.9, 1.7
2 0.8538, 2, 1.0, 0.1 0.8538, 2, 1.0, 0.1 0.8538, 2, 1.0, 0.3 0.8538, 2, 1.2, 0.7
n¼7
0.1 1.1351, 56, 224.9, 362.1 1.1351, 56, 225.8, 363.6 1.1351, 56, 245.5, 394.6 1.1351, 56, 282.7, 453.4
0.2 1.0861, 35, 87.5, 138.9 1.0861, 35, 88.3, 140.2 1.0976, 39, 107.4, 170.9 1.1102, 44, 154.2, 245.9
0.3 1.0353, 22, 35.1,54.7 1.0353, 22, 35.5, 55.5 1.0495, 25, 46.2, 72.6 1.0798, 33, 77.9, 123.4
0.5 0.9434, 10, 8.5, 12.2 0.9434, 10, 8.6, 12.4 0.9653, 12, 11.7, 17.3 1.0060, 17, 22.7, 34.9
0.7 0.8799, 6, 3.4, 4.1 0.8799, 6, 3.4, 4.1 0.8994, 7, 4.5, 5.9 0.9434, 10, 8.8, 12.6
1 0.7882, 3, 1.6, 1.3 0.7882, 3, 1.6, 1.3 0.8271, 4, 1.9, 1.8 0.8564, 5, 3.3, 4.1
1.5 0.7314, 2, 1.1, 0.3 0.7314, 2, 1.1, 0.3 0.7314, 2, 1.1, 0.5 0.7882, 3, 1.4, 1.0
2 0.7314, 2, 1.0, 0.1 0.7314, 2, 1.0, 0.1 0.7314, 2, 1.0, 0.1 0.7314, 2, 1.1, 0.4
optimal parameters ðK , L Þ ¼ (1.5828, 34) are obtained, which give (ARL1, SDRL1) ¼ (84.4,
133.9) when n ¼ 3, B ¼ 1, m ¼ 1 and g ¼ 0.1 (see Table 7). In Table 7, the precision error ratios
g 2{0, 0.1, 0.5, 1.0}, shift sizes d 2 {0.1, 0.2, 0.3, 0.5, 0.7, 1, 1.5, 2} and sample sizes n 2 {3, 5, 7}
when m ¼ B ¼ 1 are considered. It is observed in Table 7 that for fixed n, d, m (¼ 1) and B (¼
1)values, the SSGR X ~ chart is slower in detecting an out-of-control condition (a larger ARL1
value) when the value of g increases. Additionally, the SDRL1 value also increases with g. For
example, for n ¼ 5, m ¼ 1, B ¼ 1 and d ¼ 0.5, (ARL1, SDRL1) ¼ (13.3, 19.9) when there is no
measurement error (g ¼ 0) in the process, while (ARL1, SDRL1) ¼ (18.3, 27.8) when measure-
ment errors are present with g ¼ 0.5 (see Table 7). This example illustrates the negative impact of
measurement errors on the SSGR X ~ chart’s performance. It is also obvious in Table 7 that when
g ¼ 0 (no measurement error) and m ¼ B ¼ 1, the optimal parameters ðK , L Þ of the SSGR X ~
chart in Table 7 are similar to the optimal parameters (K0 , L0 ) of the SSGR X ~ chart in Table 3.
The SSGR X ~ chart reduces to the SSGR X ~ chart when g ¼ 0 and m ¼ B ¼ 1.
Table 8 shows the ARL1 and SDRL1 performances, as well as optimal parameters ðK , L Þ of
the SSGR X ~ chart for different values of B 2 {1, 2, 3, 4}, d 2 {0.1, 0.2, 0.3, 0.5, 0.7, 1, 1.5, 2} and
n 2 {3, 5, 7}, when m ¼ 1 and g ¼ 0.28. Note that g ¼ 0.28 is adopted as it is deemed as an
acceptable value of the signal-to-noise ratio (SNR) performance in attaining adequate precision
for a measurement system (Tran, Tran, and Rakitzis 2019). It is shown in Table 8 that as the
value of B increases, the ARL1 and SDRL1 values of the SSGR X ~ chart decrease. This indicates a
reduction in the negative effect of measurement errors on the performance of the SSGR X ~ chart
when B increases. For example, when n ¼ 5, m ¼ 1, g ¼ 0.28 and d ¼ 0.5, (ARL1, SDRL1) ¼
(14.8, 22.3) for B ¼ 1, while (ARL1, SDRL1) ¼ (13.4, 20.0) for B ¼ 4 (see Table 8), where both
ARL1 and SDRL1 values decrease as B increases.
Table 9 displays the ARL1 and SDRL1 values of the SSGR X ~ chart for different number of
measurements (m) per item when n, d, B (¼ 1) and g (¼ 0.28) are fixed. In Table 9, it is
COMMUNICATIONS IN STATISTICS - SIMULATION AND COMPUTATIONV
R
1673
Table 8. ðK , L , ARL1 , SDRL1 Þ values for the SSGR X chart for different values of B, d and n, where g ¼ 0.28 and m ¼ 1.
d B¼1 B¼2 B¼3 B¼4
n¼3
0.1 1.6694, 60, 292.5, 471.3 1.6694, 60, 288.9, 465.6 1.6694, 60, 288.2, 464.4 1.6694, 60, 287.9, 464.0
0.2 1.6360, 48, 169.5, 271.4 1.6296, 46, 163.7, 261.5 1.6296, 46, 162.5, 259.7 1.6296, 46, 162.1, 259.1
0.3 1.5873, 35, 89.9, 142.7 1.5828, 34, 85.2, 135.1 1.5828, 34, 84.3, 133.7 1.5781, 33, 84.0, 133.1
0.5 1.4886, 19, 27.6, 42.8 1.4796, 18, 25.6, 39.6 1.4796, 18, 25.2, 39.0 1.4796, 18, 25.1, 38.8
0.7 1.3951, 11, 10.8, 15.9 1.3951, 11, 9.9, 14.5 1.3951, 11, 9.8, 14.2 1.3951, 11, 9.7, 14.1
1 1.2853, 6, 4.0, 5.1 1.2853, 6, 3.7, 4.6 1.2853, 6, 3.6, 4.5 1.2853, 6, 3.6, 4.5
1.5 1.1511, 3, 1.6, 1.4 1.1511, 3, 1.5, 1.2 1.1511, 3, 1.5, 1.2 1.1511, 3, 1.5, 1.2
2 1.0679, 2, 1.1, 0.5 1.0679, 2, 1.1, 0.5 1.0679, 2, 1.1, 0.4 1.0679, 2, 1.1, 0.4
n¼5
0.1 1.3239, 55, 259.7, 416.6 1.3239, 55, 255.0, 409.2 1.3239, 55, 254.1, 407.8 1.3239, 55, 253.8, 407.3
0.2 1.2883, 41, 123.3, 196.6 1.2821, 39, 117.9, 187.5 1.2821, 39, 116.8, 185.9 1.2821, 39, 116.5, 185.3
0.3 1.2358, 27, 56.1, 88.5 1.2358, 27, 52.6, 82.9 1.2358, 27, 52.0, 81.9 1.2309, 26, 51.8, 81.5
0.5 1.1486, 14, 14.8, 22.3 1.1384, 13, 13.7, 20.5 1.1384, 13, 13.5, 20.1 1.1384, 13, 13.4, 20.0
0.7 1.0698, 8, 5.7, 7.7 1.0503, 7, 5.2, 7.1 1.0503, 7, 5.2, 7.0 1.0503, 7, 5.1, 6.9
1 0.9657, 4, 2.3, 2.5 0.9657, 4, 2.2, 2.2 0.9657, 4,2.1, 2.2 0.9657, 4, 2.1, 2.2
1.5 0.8538, 2, 1.2, 0.7 0.8538, 2, 1.2, 0.6 0.8538, 2, 1.2, 0.6 0.8538, 2, 1.2, 0.6
2 0.8538, 2, 1.0, 0.2 0.8538, 2, 1.0, 0.1 0.8538, 2, 1.0, 0.1 0.8538, 2, 1.0, 0.1
n¼7
0.1 1.1351, 56, 232.0, 373.4 1.1351, 56, 226.7, 365.0 1.1351, 56, 225.7, 363.4 1.1351, 56, 225.4, 362.8
0.2 1.0891, 36, 94.0, 149.3 1.0861, 35, 89.1, 141.5 1.0861, 35, 88.2, 140.1 1.0861, 35, 87.9, 139.6
0.3 1.0402, 23, 38.6, 60.4 1.0353, 22, 36.0, 56.2 1.0353, 22, 35.5, 55.4 1.0353, 22, 35.3, 55.1
0.5 0.9549, 11, 9.5, 13.7 0.9434, 10, 8.7, 12.6 0.9434, 10, 8.6, 12.4 0.9434, 10, 8.6, 12.3
0.7 0.8799, 6, 3.7, 4.7 0.8799, 6, 3.5, 4.2 0.8799, 6, 3.4, 4.1 0.8799, 6, 3.4, 4.1
1 0.7882, 3, 1.7, 1.5 0.7882, 3, 1.6, 1.4 0.7882, 3, 1.6, 1.3 0.7882, 3, 1.6, 1.3
1.5 0.7314, 2, 1.1, 0.3 0.7314, 2, 1.1, 0.3 0.7314, 2, 1.1, 0.3 0.7314, 2, 1.1, 0.3
2 0.7314, 2, 1.0, 0.1 0.7314, 2, 1.0, 0.1 0.7314, 2, 1.0, 0.1 0.7314, 2, 1.0, 0.1
observed that the ARL1 and SDRL1 values decrease with an increase in m, reflecting a positive
effect of the number of measurements (m) per item on the performance of the SSGR X ~ chart. In
addition, it is observed that for fixed values of n and d, the (ARL1, SDRL1) values of the SSGR
~ chart in Table 9 for m ¼ 10 (when measurement errors are present) are close to that of the
X
aforementioned chart without measurement error (g ¼ 0) in Table 7. For example, when n ¼ 3, g
¼ 0.28, B ¼ 1, m ¼ 10 and d ¼ 0.2, (ARL1, SDRL1) ¼ (162.4, 259.6) in Table 9, which is close to
(ARL1, SDRL1) ¼ (161.6, 258.3) for n ¼ 3, g ¼ 0, B ¼ m ¼ 1 and d ¼ 0.2 in Table 7. These findings
indicate that in the presence of small to moderate levels of measurement errors, such as g ¼ 0.28 in
Table 9, the (ARL1, SDRL1) values of the SSGR X ~ chart can be made to be close to that without
measurement error (g ¼ 0) by taking a large number of measurements per item, such as m ¼ 10.
In Table 9, it is observed that the optimal parameters (K , L ) changes when d changes for a fixed
(m, n) pair. For the case of multiple measurements (m 3), the values of ðK , L Þ are the same for a
fixed shift size d. However, for the case of a single measurement (m ¼ 1), the values of (K , L ) may
be different from that of multiple measurements (m ¼ 3, 5, 10) even though d is fixed. For example,
when n ¼ 3 and d ¼ 0.2, ðK , L Þ¼ (1.6360, 48) for m ¼ 1, which is different from the values ðK , L Þ
¼ (1.6296, 46) for m ¼ 3, 5 and 10. Note that the synthetic median chart with measurement errors
presented in Tran, Tran, and Rakitzis (2019) also shows a similar behavior.
Table 9. ðK , L , ARL1 , SDRL1 Þ values for the SSGR X chart for different values of m, d and n, where g ¼ 0.28 and B ¼ 1.
d m¼1 m¼3 m¼5 m ¼ 10
n¼3
0.1 1.6694, 60, 292.5, 471.3 1.6695, 60, 289.3, 466.2 1.6695, 60, 288.6, 465.2 1.6694, 60, 288.1, 464.4
0.2 1.6360, 48, 169.5, 271.4 1.6296, 46, 164.3, 262.5 1.6296, 46, 163.3, 260.8 1.6296, 46, 162.4, 259.6
0.3 1.5873, 35, 89.9, 142.7 1.5828, 34, 85.7, 136.0 1.5828, 34, 84.9, 134.6 1.5828, 34, 84.2, 133.6
0.5 1.4886, 19, 27.6, 42.8 1.4796, 18, 25.8, 40.0 1.4796, 18, 25.5, 39.4 1.4796, 18, 25.2, 38.9
0.7 1.3951, 11, 10.8, 15.9 1.3951, 11, 10.0, 14.6 1.3951, 11, 9.9, 14.4 1.3951, 11, 9.8, 14.2
1 1.2853, 6, 4.0, 5.1 1.2853, 6, 3.7, 4.6 1.2853, 6, 3.7, 4.5 1.2853, 6, 3.6, 4.5
1.5 1.1511, 3, 1.6, 1.4 1.1511, 3, 1.5, 1.2 1.1511, 3, 1.5, 1.2 1.1511, 3, 1.5, 1.2
2 1.0679, 2, 1.1, 0.5 1.0679,2,1.1,0.5 1.0679,2,1.1,0.5 1.0679,2,1.1,0.4
n¼5
0.1 1.3239, 55, 259.7, 416.6 1.3239, 55, 255.6, 410.1 1.3239, 55, 254.7, 408.7 1.3239, 55, 254.1, 407.7
0.2 1.2883, 41, 123.3, 196.6 1.2821, 39, 118.5, 188.5 1.2821, 39, 117.5, 186.9 1.2821, 39, 116.7, 185.7
0.3 1.2358, 27, 56.1, 88.5 1.2358, 27, 53.0, 83.5 1.2358, 27, 52.4, 82.5 1.2358, 27, 51.9, 81.8
0.5 1.1486, 14, 14.8, 22.3 1.1384, 13, 13.8, 20.7 1.1384, 13, 13.6, 20.4 1.1384, 13, 13.5, 20.1
0.7 1.0698, 8, 5.7, 7.7 1.0503, 7, 5.3, 7.2 1.0503, 7, 5.2, 7.1 1.0503, 7, 5.2, 7.0
1 0.9657, 4, 2.3, 2.5 0.9657, 4, 2.2, 2.2 0.9657, 4, 2.1, 2.2 0.9657, 4, 2.1, 2.2
1.5 0.8538, 2, 1.2, 0.7 0.8538, 2, 1.2, 0.6 0.8538, 2, 1.2, 0.6 0.8538, 2,1.2, 0.6
2 0.8538, 2, 1.0,0.2 0.8538, 2, 1.0, 0.2 0.8538, 2, 1.0, 0.1 0.8538, 2, 1.0, 0.1
n¼7
0.1 1.1351, 56, 232.0, 373.4 1.1351, 56, 227.4, 366.0 1.1351, 56, 226.4, 364.4 1.1351, 56, 225.6, 363.3
0.2 1.0891, 36, 94.0, 149.3 1.0861, 35, 89.7, 142.4 1.0861, 35, 88.8, 141.0 1.0861, 35, 88.2, 140.0
0.3 1.0402, 23, 38.6, 60.4 1.0353, 22, 36.2, 56.6 1.0353, 22, 35.8, 55.9 1.0353, 22, 35.4, 55.3
0.5 0.9549, 11, 9.5, 13.7 0.9434, 10, 8.8, 12.8 0.9434, 10, 8.7, 12.5 0.9434, 10, 8.6, 12.4
0.7 0.8799, 6, 3.7, 4.7 0.8799, 6, 3.5, 4.3 0.8799, 6, 3.4, 4.2 0.8799, 6, 3.4, 4.1
1 0.7882, 3, 1.7, 1.5 0.7882, 3, 1.6, 1.4 0.7882,3, 1.6, 1.4 0.7882, 3, 1.6, 1.3
1.5 0.7314, 2, 1.1, 0.3 0.7314, 2, 1.1, 0.3 0.7314, 2, 1.1, 0.3 0.7314, 2, 1.1, 0.3
2 0.7314, 2, 1.0, 0.1 0.7314, 2, 1.0, 0.1 0.7314, 2, 1.0, 0.1 0.7314, 2, 1.0, 0.1
obtained from the Phase-I data. Assume that a linear covariate error model with parameters
g ¼ 0.28, B ¼ 1, A ¼ 0, m ¼ 1 and n ¼ 5 is considered for the SSGR X ~ chart, where ARL0 ¼
370.4. Additionally, assume that it is important for a process mean shift of size d ¼ 0.3 to be
detected quickly. The optimal parameters K ¼ 1.2358 and L ¼ 27 are obtained from Table 8.
The control limits of the SSGR X~ chart are computed using Equations (19a) and (19b) as fol-
lows:
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
2
UCL ¼ 500:023 þ 1:2358 0:96162 þ ð0:9616 0:28Þ ¼ 501:26 (24a)
and
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
2
LCL ¼ 500:023 1:2358 0:96162 þ ð0:9616 0:28Þ ¼ 498:79: (24b)
the first sample is conforming (see Step 2 in Sec. 2). Similarly, the second sample is conforming
as X~ 2 ¼ 500.06 2 ½LCL , UCL : It is clearly seen that samples 3 to 15 are all conforming.
However, for sample 16, X ~ 16 ¼ 501.73 62 ½LCL , UCL , hence, sample 16 is non-conforming (see
Step 2 in Sec. 2). Consequently, Y1 ¼ 16 is computed. Since Y1 ¼ 16 L (¼ 27), the SSGR X ~
chart issues an out-of-control signal at sample 16 (see Step 3 in Sec. 2). Therefore, corrective
actions need to be taken to bring the out-of-control process back into the in-control condi-
tion again.
COMMUNICATIONS IN STATISTICS - SIMULATION AND COMPUTATIONV
R
1675
Figure 1. X sub-chart of the SSGR X chart, where UCL ¼ 501.26 and LCL ¼ 498.79.
8. Conclusions
This study proposes the SSGR X ~ chart to improve the efficiency of the standard X ~ chart and
investigates the statistical properties of the former via a Markov chain technique. The SSGR
~ chart is compared with the existing synthetic X,
X ~ SS synthetic X ~ and EWMA X ~ charts, in
terms of the ARL1 and SDRL1 criteria. The results show that the SSGR X ~ chart performs bet-
ter than the synthetic X ~ chart for all sizes of process mean shifts in terms of the ARL1 criter-
ion. The former also surpasses the latter in terms of the SDRL1 criterion for all shift sizes,
except when d ¼ 0.1. In comparison with the EWMA X ~ chart, the SSGR X~ chart surpasses the
former in terms of the ARL1 and SDRL1 criteria for large sizes of shifts. Furthermore, the
1676 F. N. MIM ET AL.
SSGR X ~ chart gives better ARL1 and SDRL1 performances than the SS synthetic X ~ chart, for
d 0.3 and d 0.5, respectively, except for d ¼ 2, where the two charts show equal
performances.
The SSGR X ~ chart’s performance in the presence of measurement errors is also studied. The
findings in this study show that measurement errors affect the performance of the SSGR X ~
chart. It is found that the performance of the SSGR X~ chart deteriorates as the level of measure-
ment errors increase. The negative effect of measurement errors on the SSGR X ~ chart can be
minimized by increasing the value of the constant B in the linear covariate error model or by
increasing the number of measurements per item in each sample. However, the latter approach
can increase the cost and time in process monitoring.
This study focuses on the univariate SSGR X ~ chart with measurement errors. It will be of great
interest to study the effect of measurement errors on the performance of an upgraded version of
the SSGR chart for the process mean, based on the modified side-sensitive design, also known as
the SSMGR X chart (proposed by Gadre, Joshi, and Rattihalli 2010) in the future. Moreover, the
construction of other median type group runs charts, such as the SSMGR median chart, as well
as a study on the effects of measurement errors on the said chart can be made in the future. In
addition, future research can be conducted to develop the SSGR chart for monitoring a multivari-
ate process, followed by an investigation of the chart’s performance in the presence of measure-
ment errors. As the study in this paper considers the zero-state analyses only, the steady-state
analyses can be investigated in the future.
ORCID
Faijun Nahar Mim http://orcid.org/0000-0002-4181-0523
Michael B. C. Khoo http://orcid.org/0000-0002-3245-1127
Sajal Saha http://orcid.org/0000-0002-9878-4899
Khai Wah Khaw http://orcid.org/0000-0003-2646-6477
References
Bennett, C. A. 1954. Effect of measurement error on chemical process control. Industrial Quality Control 10 (4):
17–20.
Castagliola, P., and F. O. Figueiredo. 2013. The median chart with estimated parameters. European J. of Industrial
Engineering 7 (5):594–614. doi:10.1504/EJIE.2013.057382.
Castagliola, P., P. E. Maravelakis, and F. O. Figueiredo. 2016. The EWMA median chart with estimated parameters.
IIE Transactions 48 (1):66–74. doi:10.1080/0740817X.2015.1056861.
Cheng, X. B., and F. K. Wang. 2018. The performance of EWMA median and CUSUM median control charts for
a normal process with measurement errors. Quality and Reliability Engineering International 34 (2):203–13. doi:
10.1002/qre.2248.
Costa, A. F. B., and P. Castagliola. 2011. Effect of measurement error and autocorrelation on the X chart. Journal
of Applied Statistics 38 (4):661–73. doi:10.1080/02664760903563627.
Costa, A. F. B., and M. A. G. Machado. 2013. The side-sensitive synthetic median chart. In Proceedings of the
Brazilian Symposium on Operations Research, 1301–8. Rio de Janeiro: SOBRAPO.
Gadre, M. P., K. A. Joshi, and R. N. Rattihalli. 2010. A side sensitive modified group runs control chart to detect
shifts in the process mean. Journal of Applied Statistics 37 (12):2073–87. doi:10.1080/02664760903222190.
Gadre, M. P., and V. C. Kakade. 2019. Some side sensitive group runs based control charts to detect shifts in the
process median. Communications in Statistics - Simulation and Computation. Advance online publication.
doi:10.1080/03610918.2019.1672736.
Gadre, M. P., and R. N. Rattihalli. 2004. A group runs control chart for detecting shifts in the process mean.
Economic Quality Control 19 (1):29–43. doi:10.1515/EQC.2004.29.
Gadre, M. P., and R. N. Rattihalli. 2006. Modified group runs control charts to detect increases in fraction non
conforming and shifts in the process mean. Communications in Statistics - Simulation and Computation 35 (1):
225–40. doi:10.1080/03610910500416256.
Gadre, M. P., and R. N. Rattihalli. 2007. A side sensitive group runs control chart for detecting shifts in the pro-
cess mean. Statistical Methods and Applications 16 (1):27–37. doi:10.1007/s10260-006-0013-9.
COMMUNICATIONS IN STATISTICS - SIMULATION AND COMPUTATIONV
R
1677
Haq, A. 2019. A new nonparametric synthetic EWMA control chart for monitoring process mean.
Communications in Statistics - Simulation and Computation 48 (6):1665–76. doi:10.1080/03610918.2017.1422750.
Haq, A., and M. B. C. Khoo. 2016. A new synthetic control chart for monitoring process mean using auxiliary
information. Journal of Statistical Computation and Simulation 86 (15):3068–92. doi:10.1080/00949655.2016.
1150477.
Hu, X., P. Castagliola, J. Sun, and M. B. C. Khoo. 2015. The effect of measurement errors on the synthetic X chart.
Quality and Reliability Engineering International 31 (8):1769–78.
Hu, X., P. Castagliola, X. Zhou, and A. Tang. 2019. Conditional design of the EWMA median chart with estimated
parameters. Communications in Statistics - Theory and Methods 48 (8):1871–89. doi:10.1080/03610926.2018.
1440310.
Khaw, K. W., X. Y. Chew, W. C. Yeong, and S. L. Lim. 2019. Optimal design of the synthetic control chart for
monitoring the multivariate coefficient of variation. Chemometrics and Intelligent Laboratory Systems 186:33–40.
doi:10.1016/j.chemolab.2019.02.001.
Khoo, M. B. C. 2005. A control chart based on sample median for the detection of a permanent shift in the pro-
cess mean. Quality Engineering 17 (2):243–57. doi:10.1081/QEN-200057329.
Khoo, M. B. C., H. C. Lee, Z. Wu, C. H. Chen, and P. Castagliola. 2010. A synthetic double sampling control chart
for the process mean. IIE Transactions 43 (1):23–38. doi:10.1080/0740817X.2010.491503.
Khoo, M. B. C., E. K. Tan, Z. L. Chong, and S. Haridy. 2015. Side-sensitive group runs double sampling (SSGRDS)
chart for detecting mean shifts. International Journal of Production Research 53 (15):4735–53. doi:10.1080/
00207543.2015.1043033.
Lee, M. H., and M. B. C. Khoo. 2017. Synthetic double sampling S chart. Communications in Statistics - Theory
and Methods 46 (12):5914–31. doi:10.1080/03610926.2015.1115073.
Lee, M. H., K. S. K. Y. Si, X. Y. Chew, E. M. F. Lau, and P. H. H. Then. 2019. The effect of measurement errors
on the double sampling X chart. COMPUSOFT, an International Journal of Advanced Computer Technology 8
(9):3395–401.
Linna, K. W., and W. H. Woodall. 2001. Effect of measurement error on Shewhart control charts. Journal of
Quality Technology 33 (2):213–22. doi:10.1080/00224065.2001.11980068.
Maleki, M. R., A. Amiri, and P. Castagliola. 2017. Measurement errors in statistical process monitoring: A litera-
ture review. Computers & Industrial Engineering 103:316–29. doi:10.1016/j.cie.2016.10.026.
Maravelakis, P., J. Panaretos, and S. Psarakis. 2004. EWMA chart and measurement error. Journal of Applied
Statistics 31 (4):445–55. doi:10.1080/02664760410001681738.
Mim, F. N., S. Saha, M. B. C. Khoo, and M. Khatun. 2019. A Side-sensitive modified group runs control chart
with auxiliary information to detect process mean shifts. Pertanika Journal of Science and Technology 27 (2):
847–66.
Motsepa, C. M., J.-C. Malela-Majika, P. Castagliola, and S. C. Shongwe. 2020. A side-sensitive double sampling X
monitoring scheme with estimated process parameters. Communications in Statistics - Simulation and
Computation. Advance online publication. doi:10.1080/03610918.2020.1722835.
Nguyen, H. D., Q. T. Nguyen, T. H. Nguyen, N. Balakrishnan, and K. P. Tran. 2020. The performance of the
EWMA median chart in the presence of measurement error. Artificial Intelligence Evolution 1 (1):48–62. doi:10.
37256/aie.112020401.
Park, H. I. 2009. Median control charts based on bootstrap method. Communications in Statistics - Simulation and
Computation 38 (3):558–70. doi:10.1080/03610910802585824.
Rakitzis, A. C., S. Chakraborti, S. C. Shongwe, M. A. Graham, and M. B. C. Khoo. 2019. An overview of synthetic-
type control charts: Techniques and Methodology. Quality and Reliability Engineering International 35 (7):
2081–96. doi:10.1002/qre.2491.
Riaz, A., M. Noor-Ul-Amin, M. A. Shehzad, and M. Ismail. 2019. Auxiliary information based mixed EWMA-
CUSUM mean control chart with measurement error. Iranian Journal of Science and Technology, Transactions
A: Science 43 (6):2937–49. doi:10.1007/s40995-019-00774-6.
Saha, S., M. B. C. Khoo, P. Castagliola, and A. Haq. 2021. Side-sensitive modified group runs charts with and with-
out measurement errors for monitoring the coefficient of variation. Quality and Reliability Engineering
International 37 (2):598–617. doi:10.1002/qre.2751.
Saha, S., M. B. C. Khoo, M. H. Lee, and P. Castagliola. 2018. A side-sensitive modified group runs double sampling
(SSMGRDS) control chart for detecting mean shifts. Communications in Statistics - Simulation and Computation
47 (5):1353–69. doi:10.1080/03610918.2017.1311918.
Saha, S.,. M. B. C. Khoo, P. S. Ng, and Z. L. Chong. 2019. Variable sampling interval run sum median charts with
known and estimated process parameters. Computers & Industrial Engineering 127:571–87. doi:10.1016/j.cie.
2018.10.049.
Shongwe, S. C., and J.-C. Malela-Majika. 2020. A new variable sampling size and interval synthetic and runs-rules
schemes to monitor the process mean of autocorrelated observations with measurement errors. International
Journal of Industrial Engineering Computations 11 (4):607–26. doi:10.5267/j.ijiec.2020.4.003.
1678 F. N. MIM ET AL.
Shongwe, S. C., J.-C. Malela-Majika, and P. Castagliola. 2020. The new synthetic and runs-rules schemes to moni-
tor the process mean of autocorrelated observations with measurement errors. Communications in Statistics –
Theory and Methods. Advance online publication. doi:10.1080/03610926.2020.1737125.
Tran, K. P. 2017. Run rules median control charts for monitoring process mean in manufacturing. Quality and
Reliability Engineering International 33 (8):2437–50. doi:10.1002/qre.2201.
Tran, K. P., P. Castagliola, and N. Balakrishnan. 2017. On the performance of Shewhart median chart in the pres-
ence of measurement errors. Quality and Reliability Engineering International 33 (5):1019–29. doi:10.1002/qre.
2087.
Tran, K. P., P. Castagliola, and G. Celano. 2016. The performance of the Shewhart-RZ control chart in the pres-
ence of measurement error. International Journal of Production Research 54 (24):7504–22. doi:10.1080/00207543.
2016.1198507.
Tran, P. H., K. P. Tran, and A. Rakitzis. 2019. A synthetic median control chart for monitoring the process mean
with measurement errors. Quality and Reliability Engineering International 35 (4):1100–16. doi:10.1002/qre.2447.
Wan, Q., Y. Wu, W. Zhou, and X. Chen. 2018. Economic design of an integrated adaptive synthetic chart and
maintenance management system. Communications in Statistics - Theory and Methods 47 (11):2625–42. doi:10.
1080/03610926.2016.1271425.
Wu, Z., and T. A. Spedding. 2000. A synthetic control chart for detecting small shifts in the process mean. Journal
of Quality Technology 32 (1):32–38. doi:10.1080/00224065.2000.11979969.
You, H. W., M. B. C. Khoo, P. Castagliola, and A. Haq. 2016. Monitoring the coefficient of variation using the
side sensitive group runs chart. Quality and Reliability Engineering International 32 (5):1913–27. doi:10.1002/
qre.1922.
Zhang, Y., P. Castagliola, Z. Wu, and M. B. C. Khoo, 2011. The synthetic X chart with estimated parameters. IIE
Transactions 43 (9):676–87. doi:10.1080/0740817X.2010.549547.