Research Article

(wileyonlinelibrary.com) DOI: 10.1002/qre.1284 Published online in Wiley Online Library

A New Multivariate Process Capability Index

Under Both Unilateral and Bilateral Quality
Characteristics
Majid Jalili Mahdi Bashiri and Amirhossein Amiri*
Process capability indices (PCIs) are used in statistical process control to evaluate the capability of the processes in satisfying the
customer’s needs. In the past two decades varieties of PCI are introduced by researchers to analyze the process capability with
univariate or multivariate quality characteristics. To the best of our knowledge, most famous multivariate capability indices are
proposed when the quality characteristics have both upper and lower specification limits. These indices are incapable to assess
the multivariate processes capability with unilateral specification. In this article, we propose a new multivariate PCI to analyze the
processes with one or more unilateral specification limits. This new index also accounts for all problems in the best PCIs of the
literature. The performance of the proposed index is evaluated by real cases under different situations. The results show that
the proposed index performs satisfactorily in all cases considered. Copyright © 2012 John Wiley & Sons, Ltd.

Keywords: process capability index; univariate and multivariate quality characteristic; statistical process control; one-sided
specification limit

1. Introduction

T
he development of productions of the 20th century incurs to make techniques for statistical process control improvement. These
methods improve the quality of the process results in customer satisfaction. Within the framework of statistical process control,
the process capability analysis has been received considerable attentions in the last two decades. Process capability analysis is
concerned with evaluating the capability of a process to produce products or services that meet specifications. Such capability is usually
measured by process capability indices (PCIs). A comprehensive bibliography of the literature on PCIs between years 2000 and 2009 is
given by Yum and Kim.1 PCIs can be classified in two univariate and multivariate categories.
Capability indices such as Cp, Cpk, and Cpm are typically used as measures of process capability in univariate domain. Suppose that
X1, X2, . . ., Xn denote a random sample of size n from a quality characteristic with a univariate normal distribution. If we assume USL
and LSL as the upper and lower specification limits, respectively, capability index Cp is defined as the ratio of tolerance range to
the spread of the process as follows:
USL LSL
Cp ¼ (1)
6s
Because this index cannot evaluate the location of the process (it is indifference to sample mean and only considers the deviation
of observations), Kane2 introduced Cpk for off-center processes. Kane’s index is capable to include the deviation from the target. This
index is defined as


USL m m LSL
Cpk ¼ ; (2)
3s 3s

A criticism about Cpk is that the Cpk does not account for the target of the process. For example, this index can be equal for a
process with mean on target and large variation and a process with mean far away target and small variation. In fact, it makes no clear
distinction between on-target and off-target processes. To solve this problem, Cpm as a third-generation capability index is proposed
by Chan et al.3 as follows:

Faculty of Engineering, Department of Industrial Engineering, Shahed University, Tehran, Iran

*Correspondence to: Amirhossein Amiri, Faculty of Engineering, Department of Industrial Engineering, Shahed University, Tehran, Iran.
†
E-mail: [email protected]

Copyright © 2012 John Wiley & Sons, Ltd. Qual. Reliab. Engng. Int. 2012
 M. JALILI ET AL.

USL LSL
Cpm ¼ qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi (3)
6 s2 þ ðm T Þ2

Sometimes, more than one quality characteristic describes the quality of a process. For analyzing the capability of these processes
with correlated quality characteristics, multivariate PCIs are introduced by some researchers.
Suppose that X1, X2, . . ., Xn denote a random sample vector of size n multivariate normally distributed observations of v correlated
quality characteristics for a product; that is, vector X represents the v measurements recorded from a randomly selected part. The
sample mean vector X contains the sample means of the n observations for the v quality characteristics. The sample variance–covariance
matrix S is a v  v matrix, which contains the sample variances and covariances of the n and v variate observations.P The v variate vector
X, estimates the process mean vector m, and the v  v matrix S estimates the process variance–covariance matrix . It is well known that
the contours of multivariate normal distributed quality characteristics are ellipsoids centered at vector m. This region was introduced by
Johnson and Wichern4 according to Equation (4):
X 1
ðX mÞ′ ðX mÞ⩽w2ðv;aÞ (4)

where w2ðv;aÞ is the upper 100  (1 a)% percentile of a chi-square distribution with v degrees of freedom.
According to a comprehensive survey by Wang,5 multivariate PCIs are based on following four categories:
1. The probability of nonconforming products in process with multivariate distribution (see for example Chen,6 Pal,7 Polansky8).
2. The ratio of a tolerance region to a process region (PR) such as the ones proposed by Hubele et al.,9 Taam et al.,10 Shahriari
et al.,11 Wang et al.,12 and Shahriari and Abdollahzadeh.13 Also, Pana and Lee14 introduced an index similar to the Shahriari
and Abdollahzadeh13 index.
3. Different approaches using loss functions and vector representation such as Pearn et al.15 and Abdolshah et al.16.
4. Geometric distance approach involving principle components analysis, which is originally based on the research of Wang and
Chen.17 They simplified the computation of multivariate process capability by using principal components analysis.
Other approaches such as process capability plot by Deleryd and Vannman18 and using the first four moments of nonnormal data
by Ding19 are also discussed in the literature. For studies related to PCIs, see Castagliola and Vannman,20 Noorossana,21 and Pearn and
Chen.22
Because the proposed multivariate capability index of this article belongs to the second category of multivariate PCIs, described by
Wang5, we specifically concentrate on this category and describe some PCIs of this category in more details.
Hubele et al.9 proposed a composite measure for process capability based on two quality characteristics. In the method of Taam
et al.10 (MCpm), modified tolerance is depicted based on USL and LSL of the quality characteristic. Cpm was proposed by Shahriari et al.11,
based on the original work of Hubele et al.9. This index is a three-component vector, called multivariate process capability vector.
Shahriari and Abdollahzadeh13 improved the capability index MCpm by Taam et al.10 and proposed a new multivariate process cap-
ability vector (NMPCV). When the correlation among the quality characteristics is large and/or the process is capable for some quality
characteristics and is incapable for some others, NMPCV is the most reliable tool, especially when compared with MCpm.
As explained earlier, the base of all indices in the second category is the ratio of the tolerance region to the spread of the process
region. For example, MCpm as a famous multivariate process capability is defined as

Vol:ðR1 Þ 1 MC p
MC pm ¼  ¼ (5)
Vol:ðR2 Þ D D

where R1 is a modified tolerance region and R2 is a scaled 99.73% Process region, which is an elliptical region if the underlying process
distribution is assumed to be multivariate normal. Moreover, the modified tolerance region is the largest ellipse that completely falls within
the original tolerance region with1its major axes parallel to the sides of the rectangle tolerance region. The modified Mahalanobis distance
 P 1  =2
D ¼ 1 þ ðm T Þ′ ðm T Þ is also added to this index to accounts for the distance between the process and the target. Finally, the
MCp index represents the ratio of a modified tolerance region to the process variability and is defined in Equation (6) as follows:
v 
Q v=2
ri p ½Γððv=2Þ þ 1ފ 1
i¼1
MC p ¼ P   (6)
1=2 v=2
j j p:w2ðv;aÞ Γ ðv =2 Þ þ 1ފ 1

where ri = (USLi LSLi) / 2, i = 1, 2, . . ., v and v is the number of quality characteristics.

Figure 1 illustrates the rectangular actual tolerance region, the elliptically modified tolerance region, and the process region for v = 2.
Shahriari and Abdollahzadeh13 changed the shape of tolerance region to remove some problems with the present indices intro-
duced earlier. They used the largest ellipsoid centered at target. These ellipsoid’s axes are parallel to the axes of the process ellipsoid
and are completely within the actual tolerance region. The direction of the process ellipsoid axes depends completely on the
variance–covariance structure. They defined the new modified elliptical tolerance region as follows:

Copyright © 2012 John Wiley & Sons, Ltd. Qual. Reliab. Engng. Int. 2012
M. JALILI ET AL.

Figure 1. Example of the modified tolerance region (extracted from Shahriari and Abdollahzadeh13)

P
ðx T Þ′ ðk Þ 1 ðx T Þ⩽c2 or
P 1 2 (7)
ðx T Þ′ ðx T Þ⩽kc2 ¼ c′

where k, c, and c′ are constant.

Then, they proposed the index NMCpm as

c′
NMCpm ¼ qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi  D; (8)
wð20:0027;vÞ
 P 1 =2
1
where D ¼ 1 þ ðm T Þ′ ðm TÞ and


ðUSLi Ti Þ
c′ ¼ min pffiffiffiffi ; i ¼ 1; 2; . . . ; v ; (9)
sii

where sii is the variance of ith quality characteristic.

Finally, by elimination of D from NMCpm, the new multivariate PCI was introduced as follows:

NMPCV ¼ NMCpm ; PV; LI (10)

where PV is the P value of hypothesis test H0 : m = m0 against H1 : m 6¼ m0, which is computed by using Hotelling T2 test statistic and


1 process region is completely within the tolerance region
LI ¼
0 otherwise

One of the problems with the NMPCV is that it is too sensitive to the angle between the large axis of the process region ellipsoid
and the horizontal axis. This angle completely depends on the variance–covariance structure. As an example, for two processes with equal
variation and different correlation between quality characteristics, NMCpm may compute two values with large difference (see Figures 2a
and 2b). As shown in Figures 2a and 2b, two processes with equal variation but different correlation between quality characteristics

(a) (b)

Figure 2. (a and b) Two processes with the same variation but different correlations between quality characteristics

Copyright © 2012 John Wiley & Sons, Ltd. Qual. Reliab. Engng. Int. 2012
 M. JALILI ET AL.

are shown. Because the ratio of the modified tolerance region volume to the process region volume in Figures 2a and 2b is completely
different, the value of NMCpm for the process in Figure 2b is larger than the process in Figure 2a.
Another problem with the NMPCV index is that it only can be computed for quality characteristic, which have both USL and LSL. If there
is at least one quality characteristic with unilateral specification such as LSL or USL, we were unable to compute the modified tolerance
region by using the index by Shahriari and Abdollahzadeh.13 This problem also exist for MCpm, the index proposed by Taam et al.10
In this article, a new PCI based on MCpm proposed first by Taam et al.10 is developed to compute the capability of processes, which
have at least a quality characteristic with unilateral specification limits. The new proposed index can also be used to measure the cap-
ability of processes in which all of their quality characteristics have bilateral specification limits.
The rest of the article is organized as follows. The proposed multivariate capability index is discussed in Section 2. In Section 3, a
data set from Shahriari and Abdollahzadeh13 is applied, and the performance of the proposed index is compared with the indices
proposed by Taam et al.10 and Shahriari and Abdollahzadeh.13 In addition, two real cases are applied to evaluate the performance
of the proposed index in comparison with NMPCV index. In Section 4, two real cases from the literature are applied to demonstrate
the performance of the proposed index under unilateral situation. Our concluding remarks and some future studies are given in the
final section.

2. Proposed method
As explained earlier, the basic idea of the multivariate PCIs in the second category of multivariate PCIs classified by Wang5 is the ratio
of the process region that lies in the tolerance region. The multivariate PCI of this article also uses this concept with the difference that
the PR is divided into two parts. The first part, named as conformance volume (CV), is the ratio of the process region that lies in the
modified tolerance region. The second part, which is named as nonconformance volume (NCV), is the ratio of the process region that
is out of the modified tolerance region. The proposed index is developed on the basis of the ratio of these two parts (CV and NCV),
which includes the following advantages:
1. This index considers the location and variation of a process simultaneously in the first component when both CV and NCV are
not zero rather than traditional method.
2. This index can be used for unilateral cases considering proposed the modified tolerance region for these cases.
In this method, two situations are considered. In the first situation, all of quality characteristics are bilateral, which have both USL
and LSL. In the second situation, one or more quality characteristics are unilateral, that is, they have either USL or LSL.

2.1. Bilateral situation

2.1.1. Modified tolerance and process region
In this case, the modified tolerance region is similar to Taam’s modified tolerance region, which is shown in Figure 1. Also, the distance
 P 1 1 =2
between process and target is computed as D ¼ 1 þ ðm T Þ′ ðm T Þ . Process region is also similar to Taam’s process region,
which is illustrated in Figure 1.

2.1.2. Proposed index

To propose the PCI, first we computed the modified tolerance region and the process region. Then, we considered the conformance
and nonconformance regions between the process region and the modified tolerance region.
On the basis of this concept, we proposed a new multivariate process capability based on conformance and nonconformance
regions volume (MPCNCV) as follows:
l  process region volume ðPRÞ þ conformance volume between modified tolerance region and process region ðCVÞ þ b 1
MPCNCV ¼ ; (11)
process region volume ðPRÞ þ nonconformance volume between modified tolerance region and process region ðNCVÞ D

where l is the sensitivity process capability parameter index and is explained in the Section 2.3.
Also, b is added to the numerator of the first component of the index to increase sensitivity of the index to the processes, which are
completely within the modified tolerance region. According to the simulations, it is better to apply b = 0.1, but if volume of tolerance
region is less than 1, it is advised to apply b = 0.05  volume of tolerance region because if b = 0.1, the index is overestimated.
In the other words, this number distinguishes between two processes that are completely within the modified tolerance region
with different variation. For example, consider Figure 3. Both processes A and B are completely within the modified tolerance region,
but the variation of process A is larger than the variation of process B. Without using b, the proposed index has equal value for both
processes A and B. However, using the parameter b leads to the different index value for processes A and B according to their
variations.
As shown in Equation (11), the proposed index includes two components. Generally, the PR and the modified tolerance region may
have three situations. In the first situation, the PR is completely within the modified tolerance region, in which the CV region is equal
to the PR (Figure 4a). In the second situation, the PR is completely out of modified tolerance region, so the NCV region is equal to PR
(Figure 4b).

Copyright © 2012 John Wiley & Sons, Ltd. Qual. Reliab. Engng. Int. 2012
M. JALILI ET AL.

Figure 3. Two process with different variations, which are completely within the modified tolerance region

(a) (b) (c)

Figure 4. Situation of the PR and the modified tolerance region respect to each other

In the mentioned situations, the first component of the proposed index accounts for the process variation and the second com-
ponent is the inverse of distance between the process mean and the target. Using this component, we can distinguish between
two processes with equal first component but different distance between target and process mean.
In the third situation (Figure 4c), PR lies in a part of modified tolerance region. Hence, the first component accounts for both var-
iation and distance between process mean and target. When the distance between the process mean and the target increases (with a
fixed variation), the CV region between the modified tolerance region and the PR decreases and NCV region increases. This situation
leads proposed index to be decreased. Hence, this fact leads to an increase in the sensitivity of the proposed index and the ability to
decide about the capability or incapability of a process by first component. Also, the second component can help us for better
judgment.

2.2. Unilateral situation

2.2.1. Modified tolerance and process region
To compute univariate process capability, when quality characteristic has unilateral specification, Equations (12) or (13) can be used,
USL m
Cpu ¼ (12)
3s
m LSL
Cpl ¼ (13)
3s
supposing that there is a unilateral quality characteristic with lower specification limit and the distribution of quality characteristic is
normal (see Figure 5).
As Figure 6 shows, for computing process capability, values which are greater than process mean are not considered (discarded
region is always accepted region for the process). Hence, the selected region for computing the PCI is decreased, as shown in
Figure 6.
We can use the same concept for the multivariate cases. In multivariate cases, different PCIs have been proposed by researchers as
discussed in the previous section. The special case of a multivariate PCI is a bivariate one. We used this special case to develop our
proposed multivariate capability index and discuss about different situations.
Suppose a bivariate quality characteristic with both unilateral and bilateral specification limits. The modified tolerance region in the
proposed method is depicted in Figure 7b in comparison with the bivariate quality characteristics, both have bilateral specification
limits (Figure 7a), that is, one of the quality characteristics has both USL and LSL and the other only has USL. The modified tolerance
region in Figure 7b includes two parts. Part 1 is a half tolerance region of ellipsoid which is illustrated in Figure 1. Part 2 is a rectangular

Copyright © 2012 John Wiley & Sons, Ltd. Qual. Reliab. Engng. Int. 2012
 M. JALILI ET AL.

Figure 5. A unilateral quality characteristic (with LSL)

Figure 6. Unilateral quality characteristic without values greater than the process mean

(a) (b)

Figure 7. (a) A bivariate quality characteristic with bilateral specification limits. (b) A bivariate quality characteristic with both unilateral and bilateral specification limits

tolerance region, with USL and LSL of Y1 as its two sides because Y1 has bilateral specification limits. Because Y2 is a quality character-
istic that has only an upper specification limit, the third side of the mentioned rectangular tolerance region is the mean of Y2 and the
last side is negative infinite.

Copyright © 2012 John Wiley & Sons, Ltd. Qual. Reliab. Engng. Int. 2012
M. JALILI ET AL.

Figure 8. Process region and tolerance region

2.2.2. Proposed index

Consider the PR shown in Figure 8 by color black. The PR can be located anywhere, even out of the tolerance region. However, it is
desired that the PR lies in the modified tolerance region.
To develop the PCI, we computed the CV and the NCV regions of modified tolerance and PR and considered as variation measures.
By using computed regions value (see Appendix A), the proposed index is computed as follows:

l  PR þ CV þ b 1
MPCNCV ¼ ; ′ ; (14)
PR þ NCV D

where l is a constant value and will be discussed in Section 2.3. The first component of the proposed index does not distinguish
between processes A and B, as shown in Figure 9. However, the second component, which we call location part, can identify the more
capable process. It is obvious that process B is more capable because Y2 is the smaller and better quality characteristic.
The location part can be computed by Equation (15)as follows:
1 1
¼ ð1 þ D1 Þ  (15)
D′ ð1 þ D2 Þ

where D1 is the distance between USL of Y2 and the parallel line of USL of Y2 crossing from the process mean (Figure 10). The distance
is desired to be maximized.
 
Also, D2 is the distance between the parallel line USL or LSL of Y1 crossing from USLy1 þ LSLy1 =2; USLy2 and the process mean
(Figure 10).
 
We call USLy1 þ LSLy1 =2; USLy2 as the reference point. The distance should be minimized. Note that D1 and D2 are rectilinear and
Mahalanobis distances, respectively, because D1 measures the distance of two fixed points whereas D2 measures the distance
between process mean and one fixed point.
We add 1 to both D1 and D2 in Equation (15) to avoid the index to be zero or infinite. Note that when the process mean is located
higher than the reference point (m2 ⩾ USL(Y2)), the computed PCI is wrong.
When the mean of the PR becomes far from the reference point, the second component of new index increases and leads to misleading
results. D1 in this case is considered as an undesirable distance, and the location part of the index in Equation (15) is modified as follows:

Process A Process B

Figure 9. Effect of location part on the proposed index

Copyright © 2012 John Wiley & Sons, Ltd. Qual. Reliab. Engng. Int. 2012
 M. JALILI ET AL.

(a) (b)
b

Figure 10. Role of D1 and D2 in proposed index

8
>
> 1
1 < ð1 þ D1 Þ  if m2 ⩽USLðY2 Þ
ð1 þ D2 Þ
¼ 1 1 (16)
D′ >>  if m2 ⩾USLðY2 Þ
:
ð1 þ D1 Þ ð1 þ D2 Þ

Equation (14) can be easily rewritten as a general equation for multivariate case,

l  PR þ CV þ b 1
MPCNCV ¼ ; ′ (17)
PR þ NCV D

where
1 1
¼ ka 
D′ ð1 þ Di2 Þb
(
1 þ Di1 where mi ⩽USLðYi Þ or mi ⩾LSLðYi Þ
k¼ 1 (18)
where mi ⩾USLðYi Þ or mi ⩽LSLðYi Þ
1 þ Di1
a : Number of unilateral specifications
b : Number of bilateral specifications

where Di1 and Di2 are the D1- and D2-type distance for the ith quality characteristic, respectively.

2.3. Discussion about the process capability sensitivity parameter (l)

Suppose that the first component of the proposed index value is larger than 1, which shows the process is capable. We can easily
conclude that the NCV region between the modified tolerance region and the PR should be smaller than (l/2)  PR + 0.05 (see Appen-
dix B for proof). Hence, the approximate upper bound of the NCV region is as follows:

NCV⩽l =2  PR (19)

Note that the value of l can be changed between 0 and 1. The lower the value of l, the more sensitive PCI to the NCV region. A
decrease in the l value leads to a decrease in the right-hand side of Equation (19). Hence, the NCV region has less upper bound, and
the index is sensitive to an increase in the NCV region.
Moreover, as the correlation between quality characteristics increases under a fixed variation, according to the shape of PR, the
NCV region may become larger. Hence, it is recommended one choose a small value for l to increase the sensitivity of the proposed
index to the correlated quality characteristics.
The process is capable when it is within tolerance region. Hence, the NCV region should be smaller than the tolerance region
volume, which is out of the modified tolerance region. This volume is depicted in Figure 11 by the color black.
This area is always 1 p =4 ¼ 1 0:785 ¼ 0:215 of tolerance region because

Modified tolerance region ða  b  pÞ=4 p

¼ ¼ ¼ 0:785 (20)
Tolerance region ab 4
1
We can easily conclude that this area is 0:215  0:785 ¼ 0:27 of the modified tolerance region, so we have NCV < 0.27  the mod-
ified tolerance region.

Copyright © 2012 John Wiley & Sons, Ltd. Qual. Reliab. Engng. Int. 2012
M. JALILI ET AL.

Figure 11. Tolerance region, which is not within the modified tolerance region (bilateral case)

However, according to the shape of the PR (ellipsoid shape), the NCV region can lie in half of this area when the PR is completely
within the tolerance region (in other words, two parts of this black area are used by the PR; see Figure 12b). Hence, we can conclude
NCV < (0.27/2)  the modified tolerance region.
Note that when the PR is equal to the modified tolerance region and the process is capable (see Figure 12a), NCV is less than
(0.27/2)  modified tolerance region or NCV < (0.27/2)PR.
On the other hand, when there is correlation between the quality characteristics, the PR becomes smaller than the modified tol-
erance region (see Figure 12b), which leads to a decrease in the upper bound of NCV (because PR volumes are decreased). Hence,
in this case, there is no need to decrease the value of l (as mentioned earlier), and l with a value of 0.27 is recommended.
In the unilateral case (one bilateral quality characteristic and one unilateral quality characteristic) according to the modified toler-
ance region, the value of l must be changed because there are two corners instead of four corners, so the value of l should be 0.27/
2 = 0.135 (see Figure 13).

2.4. Confidence interval for the proposed index

In this section, similar to Pearn et al.23, a confidence interval for the first component of the proposed index is developed, assuming the
sensitivity parameter of the proposed PCI (l) is equal to zero. In Section 2.3, we discussed the sensitivity parameter and recommended
the value of 0.27 for this parameter. However, as explained in the Section 2.3, when the correlation between quality characteristics

(a) (b)

Figure 12. (a) Capable process when the PR is equal to the modified tolerance region. (b) Capable process with a correlation between quality characteristics

Figure 13. Tolerance region, which is not within the modified tolerance region (unilateral case)

Copyright © 2012 John Wiley & Sons, Ltd. Qual. Reliab. Engng. Int. 2012
 M. JALILI ET AL.

increases, it is desired to use the smaller value of the sensitivity parameter. Hence, assuming l=0, the first component of the proposed
index is NCVþb
PRþCV . Note that PR volume follows a chi-square distribution with v degrees of freedom. Hence, a 100(1 a)% confidence
interval for the first component of MPCNCV can be established as follows:


NCV þ b
p L⩽ ⩽U ¼ 1 a

PR þ CV 
L 1 U
p ⩽ ⩽ ¼1 a

NCV þ b PR þ CV NCV þ b
NCV þ b NCV þ b
p ⩽PR þ CV⩽ ¼1 a

U L 
NCV þ b NCV þ b
p CV⩽PR⩽ CV ¼ 1 a
U L
PR e w2n;n

As a result, we have


NCV þ b NCV þb
p CV⩽w2n;v ⩽ CV ¼ 1 a
U L

R NCVþb CV
Let y ¼ w2n;v , then L
NCVþb fY ðy Þdy ¼ 1 a Hence,
U CV

NCV þ b
FY 1 ða=2Þ ¼ CV
U
and
NCV þ b
FY 1 ð1 a=2Þ ¼ CV
L
Rz
where FY ðz Þ ¼ 0 fY ðy Þdy; thus, a 100(1 a)% confidence interval for the first component of MPCNCV can be obtained as
1
FY ð1 a=2Þ þ CV FY 1 ða=2Þ þ CV
;
NCV þ b NCV þ b

3. Numerical examples for bilateral situation

We investigated the performance of the proposed index through three examples. The first example is extracted from Shahriari and
Abdollahzadeh,13 the second and third ones are real cases.

3.1. Example 1
In the first data set, six processes with bivariate quality characteristics, distributed as bivariate normal, are considered. In this data set,
it is assumed that both quality characteristics are bilateral and the process mean is on target. The process variance–covariance matrix,
the process mean, the tolerance limits for each quality characteristic, and the coefficient of correlation between characteristics are
given in Table I.
In Table II, results of the applying proposed index, NMPCV and MCpm, are given.
The results show that the problems of index by Taam et al.10 reported by Shahriari and Abdollahzadeh13 have been considered in
the proposed index. Hence, the proposed index does not lead to misleading results for correlated data sets. Figure 14 (panels 1–6)
contains graphical representation of each process for the first example.
Note that the first component of NMPCV only provides a comparison of the volume of the regions (variation of process) and does
not account for distance between process mean and target. This distance is measured by the second and third components of
NMPCV. Hence, for example, it is possible that a process has less variation but more distance between mean and target rather than
other process. Consequently, it leads to opposite components of PCI for processes. Hence, Decision Maker (DM) is confused in deciding
which process is more capable, although the first component of proposed index accounts for variation and distance between process
mean and target simultaneously. Hence, when the process mean is not on target, the proposed index gives a better vision to decision
maker for judgment about process capability.
Figure 15 illustrates the results of NMPCV and the proposed index for the first example when the process mean is on the target.
The trends of the two figures are similar, and it shows that the proposed index leads to the same results when the process mean is on
the target.

Copyright © 2012 John Wiley & Sons, Ltd. Qual. Reliab. Engng. Int. 2012
M. JALILI ET AL.

Table I. The characteristics of processes for example 1

Coefficient
Variance– Process Tolerance of
Process Correlation covariance matrix mean limits correlation
1 Weak 0:7115 0:3035 m = (5.5, 5.5) (2.5, 8.5) 0.38
0:3035 0:7335
2 Moderate 0:8349 0:4448 m = (5.5,5.5) (2.5, 8.5) 0.51
0:4448 0:9178
3 High 0:5483 0:6337 m = (5.5,5.5) (2.5, 8.5) 0.88
0:6337 0:8761
4 Very high 0:8579 0:9385 m = (5.5,5.5) (2.5, 8.5) 0.99
0:9385 1:0441
5 High, with the process very capable for 0:3 0:6 m = (5.5,5.5) (2.5, 8.5) 0.81
characteristic 1 and incapable for characteristic 2 0:6 1:8
6 Very high, with the process very capable for 0:4 0:65 m = (5.5,5.5) (2.5, 8.5) 0.94
characteristic 1 and incapable for characteristic 2 0:65 1:2

Table II. The comparison of PCI for the first example

Process MCpm (Taam et al.10) NMPCV (Shahriari and Abdollahzadeh13) MPCNCV (proposed index)
1 1.16 [1.01, 1, 1] [1.106, 1]
2 1 [0.91, 1, 0] [0.971, 2]
3 2.29 [0.93, 1, 0] [0.967, 1]
4 6.22 [0.825, 1, 0] [0.832, 1]
5 1.80 [0.6501, 1, 0] [0.819, 1]
6 3.17 [0.796, 1, 0] [0.912, 1]

3.2. Example 2
Taam et al.10 discuss about a process in an integrated circuit manufacturing in which two correlated quality characteristics, width and
thickness, follows a bivariate normal distribution with target values of 4.5 and 0.75 mm, respectively. The sample mean and the sample
 
P 0:02 0:009
variance–covariance matrix are x ¼ ð4:3; 0:8Þ and ¼ , respectively. The specification limits for width and thickness
0:009 0:006
are [0.5, 1] and [4, 5], respectively. Figure 16 depicts process and tolerance regions for this real case.
The value of the proposed index for this example is [0.88, 0.28], which shows that the process is not capable (note that b = 0.05 
tolerance region). Table III compares the result of the proposed index with Taam et al. 10 and the methods of Shahriari and
Abdollahzadeh.13

3.3. Example 3
Another example is extracted from Sultan24 (see Figure 17). This example is used by many researchers such as Chen,6 Shahriari and
Abdollahzadeh,13 and Pana and Lee.14
The tolerance region is a rectangle with target values T1 = 177 for H, T2 = 53 for S, and specification limits (112.7, 241.3) and (32.7,
73.3) for H and S, respectively. Using the data set, we calculated the sample mean vector X ¼ ½177:2; 52:32Š and the sample variance–
337:8 85:3308
covariance matrix s ¼ .
85:3308 33:6247
As depicted in Figure 16, the PR slightly exceeds the boundaries of actual tolerance region (LSL of H). Hence, the process capability
should be less than 1. As shown in Table IV, MPCNCV measures the capability correctly because the first component of MPCNCV is 0.92,
which is less than 1. NMPCV index shows this fact with the third component as it is equal to 0.
As mentioned earlier, the other advantage of the proposed index is that it considers both variation and distance between
process mean and target in the first component. However, when the first component of the proposed index for two processes
is equal, the second component as the distance between target and process mean can be used to find the more capable
process.

Copyright © 2012 John Wiley & Sons, Ltd. Qual. Reliab. Engng. Int. 2012
 M. JALILI ET AL.

(1) (2)

(3) (4)

(5) (6)

Figure 14. Graphical representation of the six processes for the first example

1.2

0.8

0.6 NMPCV
MPC (NCV)
0.4

0.2

0
1 2 3 4 5 6

Figure 15. Values of NMPCV and MPCNCV for the first example

Copyright © 2012 John Wiley & Sons, Ltd. Qual. Reliab. Engng. Int. 2012
M. JALILI ET AL.

Figure 16. Process and tolerance region for example 2

Table III. Comparison among three PCIs for example 2

MCpm (Taam et al.10) NMPCV (Shahriari and Abdollahzadeh13) MPCNCV (proposed index)

0.464 [0.94, 0.002, 0] [0.88, 0.28]

Figure 17. Process and tolerance regions for example 3

Table IV. Comparison among PCIs for Example 3

MCpm (Taam et al.10) NMPCV (Shahriari and Abdollahzadeh13) MPCNCV (proposed index)

1.82 [1.01, 0.6, 0] [0.92, 0.97]

4. Numerical examples for unilateral situation

In this section, we investigate the performance of the proposed index in the unilateral cases through two examples which are real
cases.

4.1. Example 4
A biological reduction of the ethyl 4-chloro acetoacetate process for the production of optically pure compound is applied to demon-
strate the effectiveness of the proposed index.25,26 S-4-chloro-3-hydroxybutyric acid ethyl ester (or S-CHBE) can be used to synthesize
various optically active compounds such as antihypertensive drugs and antibiotics. The compound can be produced by adding ethyl
4-chloro acetoacetate to baker’s yeast. The reaction produces S-CHBE with a small amount of enzymes in the yeast cell.

Table V. The results of the experiments for example 4

Experiment Yield% EE% MPCNCV (proposed index)
1 72.10 69.50 69.90 88.63 83.88 88.84 [1.06796, 2213.7]
69.60 76.10 71.80 85.63 89.22 83.29
2 72.80 78.20 67.90 80.40 74.42 82.23 [1.06786, 2085.7]
76.03 66.30 77.30 78.34 80.59 78.01

Copyright © 2012 John Wiley & Sons, Ltd. Qual. Reliab. Engng. Int. 2012
 M. JALILI ET AL.

In this real case, the yield and enantiomeric excess (EE) are determined as two quality characteristics that are the larger-the-better
type. Pertinent literature on this reaction indicates that the yield lies in 42% to 62% and the EE lies in 15% to 85% for S-CHBE. Table V
summarizes the results of the two confirmed experiments for yield and EE as well as the values of the proposed PCI.
The results of the index in Table V considering Figures , 18a–18c show the effectiveness of the proposed index in measuring the
capability of the processes. Note that because two processes are completely in the modified tolerance region, the first component of
proposed index only consider the variation of processes and distance from target is measured with the second component. To com-
pute the proposed index, l = 0. 27/4 = 0. 0675 is used because two quality characteristics are unilateral, and there is one corner in the
modified tolerance region.

(a)

(b)

(c)

Figure 18. (a) Relationship between the tolerance region and the modified tolerance region for example 4. (b–c) PR and the modified tolerance region for processes of
example 4

Copyright © 2012 John Wiley & Sons, Ltd. Qual. Reliab. Engng. Int. 2012
M. JALILI ET AL.

Figure 19. PR and the modified tolerance region for example 5

4.2. Example 5
This example is extracted from Maghsoodloo and Huang.27 A random sample of size 20 plastic pipes is taken from a bivariate normal
process, where Y1 is the strength (psi) and Y2 is the total weight per units (lb). The specifications limits are [18.5 , 1] for Y1 and
[1.875 , 2.125] for Y2, and the target of Y2 is 2. Suppose the process mean for Y1 is 20. The data sets are as follows:

ð18:60; 1:92Þ; ð19:51; 2:13Þ; ð20:84; 2:10Þ; ð20:63; 2:08Þ; ð19:84; 1:99Þ;
ð19:07; 1:93Þ; ð19:19; 1:98Þ; ð19:86; 2:06Þ; ð20:99; 2:03Þ; ð18:33; 1:86Þ;
ð19:43; 1:97Þ; ð19:19; 2:01Þ; ð20:33; 2:12Þ; ð20:77; 2:10Þ; ð20:12; 1:95Þ;
ð18:74; 1:91Þ; ð19:51; 2:06Þ; ð20:09; 2:04Þ; ð17:70; 1:96Þ; ð19:67; 2:01Þ

Figure 19 depicts the modified tolerance and PRs. The computed proposed index is equal to [0.324, 106.7], which shows that pro-
cess is not capable. This result can be easily confirmed by the Figure 19.

5. Conclusions
In this article, a new multivariate PCI based on the NCV region (MPCNCV) for processes with multivariate normal distribution was intro-
duced. Comparison studies showed that it can dominate other existing approaches and solve their problems. Moreover, the proposed
index can be used for unilateral cases. The effectiveness of the proposed index is evaluated through different real cases and numerical
examples. In addition, a confidence interval for the proposed index under an assumption on sensitivity parameter was developed.
Investigating the inferential properties of the proposed index deeply can be considered as a future research. Also, the PCI for pro-
cesses with other multivariate distributions rather than normal can be considered by researchers.

Appendix A: Computing CV and NCV regions as well as PR area

To compute CV and NCV regions for the proposed index, we used integral. Consider a unilateral process, which is depicted in
Figure A1.
First, we divided CV region in two parts from process mean of Y2 (Figure A1).

Figure A1. Tolerance and PRs for a unilateral process

Copyright © 2012 John Wiley & Sons, Ltd. Qual. Reliab. Engng. Int. 2012
 M. JALILI ET AL.

a b
Suppose S-1 matrix defined as . Equations (A1) and (A2) denote equations of the modified tolerance region and PR,
c d
respectively,
 2  2
y1 my1 y2 my2
 2 þ  2 ¼ 1 (A1)
USLy1 my1 USLy2 my2

aðy1 y 1 Þ2 þ d ðy2 y 2 Þ2 þ ðb þ cÞðy1 y 1 Þðy2 y 2 Þ w2ðv;aÞ ¼ 0 (A2)

To compute the area of part 1, the first area of the black region is computed with integral as follows:
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Z B Z ðUSLy my Þ2 ðUSLy my Þ2 y1 2
2 2 1 1
p¼ qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 1 dy2 dy1 (A3)
w2 2 2
ðv;aÞ aðy1 y1 Þ ðbþcÞ ðy1 y 1 Þ ðbþcÞðy1 y 1 Þ
A þ þy 2
d d 4d 2 2d

Consequently, the area of part 1 is computed as follows:

  
Part 1 ¼ USLy1 my1 USLy2 my2 p p (A4)

Then part 2 is computed using Equation (A5),

0 sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 1
Z D w2ðv;aÞ aðy1 y 1 Þ ðb þ cÞ2 ðy1 y 1 Þ2
@ ðb þ cÞðy1 y 1 Þ
Part 2 ¼ þ þ y 2 A dx (A5)
C d d 4d2 2d

CV region is equal to part 1 + part 2.

Finally, the NCV region is computed by the subtraction of the CV region from the PR volume.
To compute the PR volume, we need to have the value of large and small diameters of PR. Assume PR function in Equation (A2) is
rewritten as follows:

Aðy1 y 1 Þ2 þ B ðy1 y 1 Þðy2 y 2 Þ þ C ð y2 y 2 Þ2 ¼ 1 ; (A6)

where A, B, and C are constant parameters. θ as smaller angle between large or small diameter with horizontal axis is computed as
follows:
1 1 B
θ¼ tan (A7)
2 A C

Moreover, define a′ and b′ as follows:

a′ ¼ cosðθÞ (A8)

b′ ¼ sinðθÞ (A9)

Then large and small diameters are computed as follows:

( )
1 1
Large diameter ¼ 2  max pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ; pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi (A10)
2 2 2 2
a′ A þ a′ b′ B þ Cb′ b′ A a′ b′ B þ a′ C
( )
1 1
Small diameter ¼ 2  min pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ; pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi (A11)
2 2 2 2
a′ A þ a′ b′ B þ Cb′ b′ A a′ b′ B þ a′ C

Finally, PR volume is computed with Equation (A12),

Process region volume ðPRÞ ¼ ðLarge diameter  Small diameter  pÞ=4 (A12)

Appendix B: The upper bound of the NCV region for capable processes
If the process is capable, the PCI should be greater than one; consequently, we have (b = 0.1):

Copyright © 2012 John Wiley & Sons, Ltd. Qual. Reliab. Engng. Int. 2012
M. JALILI ET AL.

l  PR þ CV þ 0:1
⩾1 (B1)
PR þ NCV
CV ¼ PR NCV (B2)

Substituting Equation (B2) in Equation (B3), we have

NCV⩽l =2  PR þ 0:05 (B3)

Acknowledgements
The authors are grateful to the anonymous reviewer for helpful and constructive comments, which led to improvement in the article.
Also we wish to thank Mohammad Reza Abdollahzadeh for explaining his article let us to compare our proposed method with his
method and Mahdi Rafie and Mostafa Abbasi for their helps on mathematical issues.

References
1. Yum BJ, Kim KW. A bibliography of the literature on process capability indices: 2000–2009. Quality and Reliability Engineering International 2010;
27(3):251–268.
2. Kane VE. Process capability indices. Journal of Quality Technology 1986; 18(1):41–52.
3. Chan LK, Cheng SW, Spiring FA. A new measure of process capability: Cpm. Journal of Quality Technology 1988; 20(3):162–175.
4. Johnson RA, Wichern DN. Applied multivariate statistical analysis. Prentice Hall: Englewood Cliffs, NJ, 1992.
5. Wang FK. Quality evaluation of a manufactured product with multiple characteristics. Quality and Reliability Engineering International 2006; 22(2):
225–236.
6. Chen HF. A multivariate process capability index over a rectangular solid tolerance zone. Statistica Sinica 1994; 4(2):749–758.
7. Pal S. Performance evaluation of a bivariate normal process. Quality Engineering 1999; 11(3):379–386.
8. Polansky AM. A smooth nonparametric approach to multivariate process capability. Technometrics 2001; 43(2):199–211.
9. Hubele NF, Shahriari H, Cheng CS. A bivariate process of capability vector. In Statistical Process Control in Manufacturing, Keats JB, Montgomery DC
(eds.). Marcel Dekker: New York, NY, 1991; 299–310.
10. Taam W, Subbaiah P, Liddy JW. A note on multivariate process capability indices. Journal of Applied Statistics 1993; 20(3):339–351.
11. Shahriari H, Hubele NF, Lawrence FP. A multivariate process capability vector. Proceeding of the 4th Industrial Engineering Research Conference. Institute of
Industrial Engineers 1995; 304–309.
12. Wang FK, Hubele NF, Lawrence FP, Miskulin JD, Shahriari H. Comparison of three multivariate process capability indices. Journal of Quality Technol-
ogy 2000; 32(3):263–275.
13. Shahriari H. Abdollahzadeh M. A new multivariate process capability vector. Quality Engineering 2009; 21(3):290–299.
14. Pana JN, Lee CY. New capability indices for evaluating the performance of multivariate manufacturing processes. Quality and Reliability Engineering
International 2010; 26(1):3–15.
15. Pearn WL, Kotz S, Johnson NL. Distributional and inferential properties of process capability indices. Journal of Quality Technology 1992; 24(4):216–231.
16. Abdolshah M, Yusuff RM, Tang S, Ismail YB. New process capability index using taguchi lost functions. Journal of Applied Science 2009; 9(20):3775–3779.
17. Wang FK, Chen JC. Capability index using principal component analysis. Quality Engineering 1999; 11(1):21–27.
18. Deleryd D, Vannman K. Process capability plots-a quality improvement tool. Quality and Reliability Engineering International 1999; 15(3):213–227.
19. Ding J. A method of estimating the process capability index from the first four moments of non-normal data. Quality and Reliability Engineering
International 2004; 20(8):787–805.
20. Castagliola P, Vannman K. Monitoring capability indices using an EWMA approach. Quality and Reliability Engineering International 2007; 23(7):769–790.
21. Noorossana R. Process capability analysis in the presence of autocorrelation. Quality and Reliability Engineering International 2002; 18(1):75–77.
22. Pearn WL, Chen KS. Estimating process capability indices for non-normal pearsonian populations. Quality and Reliability Engineering International
1995; 11(5):386–388.
23. Pearn, WL, Wang FK, Yen CH. Multivariate capability indices: distributional and inferential properties. Journal of Applied Statistics 2007; 34(8):941–962.
24. Sultan TL. An acceptance chart for raw materials of two correlated properties. Quality Assurance 1986; 12:70–72.
25. Tong LI, Wang CH, Houng JY, Chen JY. Optimizing dynamic multiresponse problems using the Dual-Response-surface method. Quality Engineering
2001; 14(1):115–125.
26. Hsieh KL, Tong LI, Chiu HP, Yeh HY. Optimization of a multi-response problem in Taguchi’s dynamic system. Computers and Industrial Engineering
2005; 49(4):556–571.
27. Maghsoodloo S, Huang LH. Quality loss functions and performance measure for a mixed bivariate response. Journal of Manufacturing Systems 2001;
20(2):73–88.

Authors' biographies
Majid Jalili is a master student in Industrial Engineering at Shahed University. He received his Bachelor of Science in Industrial Engi-
neering degree from Qom University in 2009. His research interests include statistical process control and design of experiments.
Mahdi Bashiri is an assistant professor at Shahed University. He holds a Bachelor of Science in Industrial Engineering degree from Iran
University of Science and Technology and MS and PhD degrees in this field from Tarbiat Modares University. His research interests are
statistical quality control, design of experiments, and multicriteria decision making.
Amirhossein Amiri is an assistant professor at Shahed University. He holds BS, MS, and PhD degrees in Industrial Engineering from
Khajeh Nasir University of Technology, Iran University of Science and Technology, and Tarbiat Modares University, respectively. He is a
member of the Iranian Statistical Association. His research interests are statistical quality control, profile monitoring, and Six Sigma.

Copyright © 2012 John Wiley & Sons, Ltd. Qual. Reliab. Engng. Int. 2012