METRON - International Journal of Statistics

2007, vol. LXV, n. 3, pp. 305-318

LINGYUN ZHANG – K. GOVINDARAJU

On probability limits
for phase II S control chart

Summary - The Shewhart S charts are based on 3 sigma limits. The control limit
constant 3 is due to the normal approximation to the underlying distribution namely
the chi distribution, which is rather skewed. In this paper, we critically review the 3-
sigma control limits used for Phase II S charts; we examine the ARL properties. Based
on our study, we recommend the adoption of either of the equally well performing
ARL-Unbiased design or the Two-Point design for Phase II S charts in order to detect
both quality improvement and degradation.

Key Words - S charts; ARL-unbiased design; Quality control.

1. Introduction

Quality is a basically latent variable and can be measured in many ways.

One of the operational ways of measuring quality is to employ a measure of
variability in the quality characteristic of interest. Montgomery (2005) interprets
quality as one inversely proportional to variability and then defines quality
improvement as the reduction of variability in processes and products. Control
charts such as the X̄ charts are useful only to detect the changes in the process
mean level whereas S and R charts monitor the process variability. Hence
quality improvement is often detectable by the S or R chart. Unfortunately the
traditional approach to setting up control limits for these charts is based on the
ad hoc three sigma limits. This often results in an inappropriate lower control
limit, greatly diminishing the ability of the chart to detect quality improvement.
In this paper, we improve the design of S chart resulting in better control limits
to detect quality improvement as well as adequately sensitive to detect quality
deterioration. Monitoring both quality deterioration and improvement using the
same chart is not a new idea. The new control charting procedures developed in
the recent years emphasize the detection of quality improvement. For example,
see Acosta-Mejia (1998) for a new CUSUM chart based on the subgroup range

Received June 2007 and revised October 2007.

306 LINGYUN ZHANG – K. GOVINDARAJU

which will help to search for special causes that may result in a smaller process
spread. Chen et al. (2001) presented a new single EWMA-type chart based
on the transformations of sample mean and sample variance in place of two
separate EWMA charts for the mean and variance.
In the S chart, the control statistic plotted is the subgroup standard devia-
tion, i.e. at time i we plot
,
-
- 1  n
Si = . (xi j − x̄i )2 , for i ≥ 1 ,
n − 1 j=1

where {xi1 , xi2 , . . . , xin } denotes a subgroup of size n and its mean is x̄i . We
assume that all observations within and between subgroups are independent and
are from a normal population. The distribution of Si is closely related to a chi
(or chi-square) distribution, which is a right skewed distribution, particularly
for small subgroup sizes. The following lower and upper control limits recom-
mended in textbooks (see for example Montgomery 2005, Devore and Farnum
2005) are based on the 3 sigma limits, derived using a normal approximation
to the chi distribution.

LC L = c4 σ0 − 3σ0 1 − c42 , (1)

U C L = c4 σ0 + 3σ0 1 − c42 , (2)

where c4 = (2/(n − 1))1/2 (n/2)/ ((n − 1)/2) and σ0 , which is known, is

the true or well estimated process standard deviation. The above limits are for
Phase II or current production monitoring, as against the Phase I assessment
of control for retrospective data (for further discussion on Phase I and Phase
II monitoring, see Woodall and Montgomery (1999)).
In the next section, we justify why 3-sigma limits should not be used for
Phase II S charts. We then show how to set up control limits according to the
following Average-Run-Length (ARL) criteria:
(C1) When the process standard deviation is σ0 , the ARL of the S chart is
equal to (or not less than) some prefixed value, e.g. 370, and
(C2) If the process standard deviation has changed to another value, the ARL
should be short.
Note that Zhang et al. (2005) has used the essentially same criteria to
discuss the statistical design of S 2 charts. However it should be noted that
S charts are popular at the shop floor than S 2 charts mainly because the
standard deviation preserves the units of measurements as the original quality
characteristic and an engineering meaning can be given for various levels of
the process sigma. The earlier work on S 2 charts of Zhang et al. (2005) does
 On probability limits for phase II S control chart 307

not specifically address the issue of Phase II control, whereas the current work
does.
We denote throughout this paper that the process standard deviation is
σ = ρσ0 , where ρ > 1, 0 < ρ < 1 and ρ = 1 indicate whether the process
standard deviation has increased, decreased and is in control, respectively. The
average run length ARL(ρ) is therefore a function of ρ and other parameters
brought in by the lower and upper control limits. Suppose that LC L and
U C L have been specified, say LC L = aσ0 and U C L = bσ0 . It is shown in
Appendix A that

1
ARL(ρ) = # $, (3)
1 − F(sb2 /ρ 2 ) − F(sa 2 /ρ 2 )
x
where s = (n − 1)/2 and F(x) = 0 (1/ (s)) u s−1 e−u du.

2. Why to avoid three-sigma limits?

We assert that the following are the reasons against using the 3-sigma
limits given as equations (1) and (2):
1) The lower control limit LC L must be positive to detect quality improve-
ment. This forces the subgroup size n to be greater than or equal to 6.
2) The S chart with the 3-sigma limits does not achieve the advertised in-
control ARL such as 370.
3) The S chart with the 3-sigma limits is not sensitive to decreases in process
standard deviation. This means that the important objective of detecting
quality improvement will not be immediately realised.
4) The problems listed in 2) and 3) becomes less severe only when n is as
large as 20. However, in practice n is usually small, with n = 5 and
n = 10 being the reasonable choices.
Table 1: Some ARL values of 3-sigma S charts with different subgroup size n. The in-control
ARL equals ARL(1.0).

n ARL(0.5) ARL(0.75) ARL(0.9) ARL(1.0) ARL(1.1) ARL(1.25) ARL(1.5)

6 5054.6 23184.4 1052.8 198.5 59.8 17.4 5.3

7 412.8 4076.2 1250.0 221.8 61.2 16.4 4.7
8 113.2 1492.0 1296.6 239.9 61.6 15.4 4.2
9 47.0 768.6 1236.4 254.1 61.3 14.5 3.9
10 24.3 468.9 1128.8 265.6 60.7 13.6 3.5
15 3.8 105.2 647.0 300.2 55.2 10.1 2.5
20 1.7 43.4 410.7 317.6 49.3 7.8 1.9
308 LINGYUN ZHANG – K. GOVINDARAJU

It is straightforward to prove the first assertion given above. The evidence

for assertions 2), 3) and 4) is given in Table 1 and Figure 1. We see from
Table 1 that all the in-control ARL values are less than 370. On the other
hand, if there is a small improvement in the process quality by the reduc-
tion in process standard deviation, then it will take a longer time to detect
(see ARL(0.9) values). If the process standard deviation decreases by 25 %,
then more subgroups must be plotted to detect an improvement (see values of
ARL(0.75) for n = 6(1)10)) when compared to the false alarm occurrence (i.e.
ARL(0.75) > ARL(1) for n = 6(1)10). We see also from Figure 1 that the
peaks of ARL curves are not at ρ = 1 but at ρ < 1; this phenomenon is called
ARL biased in the literature. The bias in ARL explains that the chart is not
sensitive to detecting a drop in process standard deviation.
10

n=6
n=7
n=8
8

n=9
n=10
n=15
n=20
ln(ARL(rho))
6
4
2
0

0.0 0.5 1.0 1.5 2.0

rho
Figure 1. Plotting of 3-sigma S chart’s ARL (on natural logarithmic scale) against rho for various n.
The in-control ARL equals ARL(1.0).

3. Probability limits

The study given in the last section convinced us that 3-sigma limits should
not be used for Phase II S charts. It is desirable to obtain the control limits
meeting the criteria (C1) and (C2) mentioned in the Introduction section. The
 On probability limits for phase II S control chart 309

criterion (C1) is equivalent to setting that the false alarm rate is equal to (or
less than) a preassigned small fraction α (e.g. α = 0.0027). To be precise,
(C1) requires that the sum of left and right tail probability is equal to α when
process standard deviation σ = σ0 . Denote the left and right tail probability
by αl and αr , respectively, then (C1) is translated to
αl + αr = α . (4)
Given the tail probabilities, we obtain the control limits as the probability limits
as: 
/
χα2l , n−1 c1
LC L = σ0 ≡ σ0 ,
n−1 n−1

/
χ1−α
2
r , n−1
c2
U C L = σ0 ≡ σ0 ,
n−1 n−1
where χα2l , n−1 and χ1−α
2
r , n−1
, denoted by c1 and c2 , respectively, are the αl and
1 − αr quantiles of the chi-square distribution with n − 1 degrees of freedom.
Note that once αl is fixed, then αr is determined from (4) and then the above
probability limits LC L and U C L can be obtained. Now, the important question
is: How do we determine αl (or c1 )?
3.1. Half-Half split design

One simple solution to the above problem is the half-half split, i.e. let
αl = αr = α/2, which is the conventionally so-called probability limits design
(see Ryan 2000). However, the underlying chi (or chi-square) distribution is
skewed and this design of the S chart ignores criterion (C2); hence this is not
the best design.
3.2. ARL-Unbiased Design

An alternative to the Half-Half Split Design is the ARL-Unbiased Design,

where αl is chosen such that the ARL function attains its maximum at ρ = 1.
Some authors have considered ARL-Unbiased design for S charts, see for
example Champ and Lowry (1994) and Pignatiello et al. (1995). The following
is our results on ARL-Unbiased design for S charts. Our approach is different
from that in Champ and Lowry (1994), and our results are more accurate than
those given by Pignatiello et al. (1995).
Recall that c1 = χα2l , n−1 and c2 = χ1−α 2
r , n−1
. The c1 and c2 that gives an
ARL-Unbiased design must satisfy the following two equations:
  c2 /2 1

 u s−1 e−u du = 1 − α
c1 /2 (s)
(5)

 c1 − c2
= 1.
2s(ln(c1 ) − ln(c2 ))
310 LINGYUN ZHANG – K. GOVINDARAJU

The derivation of (5) is given in Appendix B, and an R program, which allows

us to obtain ARL-Unbiased designs for any specified values of n and α, is
given in Appendix D.

3.3. Two Points Criterion Design

Criterion (C1) is indeed a restricted design, but criterion (C2) suggests that
we should seek an “optimal” αl such that the resulted S chart has uniformly
small out-of-control ARL values. Unfortunately, such an optimal αl does not
exist as pointed out by the following Theorem.
Theorem. Suppose we have two choices for αl in Equation (4), say αl1 and αl2 .
Let ARL(ρ; αli ) denote the ARL function when we choose the left tail probability
to be αli , where i = 1, 2. If αl1 < αl2 , then

ARL(ρ; αl1 ) > ARL(ρ; αl2 ), for 0 < ρ < 1 ,

and
ARL(ρ; αl1 ) < ARL(ρ; αl2 ), for ρ > 1 .
A proof of the above theorem is given in Appendix C. This theorem shows
that with the restriction of (4) it is impossible to find a left tail probability αl
such that the resulted S chart has uniformly smaller out-of-control ARL than
any other S chart. In fact, choosing a small αl makes the chart sensitive to
increases in process standard deviation, while choosing a large αl makes the
chart sensitive to decreases in process standard deviation. Therefore, we need
to replace criterion (C2) by a weaker criterion. In line with Zhang et al. (2004,
2005), we use the following one.
(C2 ) The best choice of αl is determined by minimizing

ARL(1 − ε1 ; αl ) + ARL(1 + ε2 ; αl ) ,

with αl restricted by (4).

In (C2 ), parameters ε1 and ε2 are percentage decrease and increase, re-
spectively, from the target process standard deviation σ0 that are considered
practically significant for detection. The values of ε1 and ε2 must be specified
by practitioners. It might be sensible to have ε1 = ε2 = 0.5, for example. If
monitoring deterioration is more important than monitoring quality improve-
ment, we can set ε2 < ε1 . The idea behind (C2 ) is that we are trying to
pull down the ARL curve from two chosen points, (1 − ε1 , ARL(1 − ε1 )) and
(1 + ε2 , ARL(1 + ε2 )); for this reason, this design is called the two points
criterion (TPC) design.
 On probability limits for phase II S control chart 311

When ε1 and ε2 are given, the objective function in (C2 ) is a function of

αl , which because of (4) is constrained in the range from 0 to α. We have
done some numerical study on the objective function in (C2 ). The study makes
us believe that a unique solution to TPC design always exists. However, it is
possible that the optimal αl takes on 0 or α (the boundary values); this happens
when either pull-position (i.e. 1 − ε1 or 1 + ε2 ) is relatively far away from 1
and that makes the “pull-down force” from one side stop working.
An R program, which allows us to obtain TPC designs for any specified
values of n, α, ε1 and ε2 , is provided in Appendix D.

4. Comparisons

For n = 5 or n = 10 and α = 0.00027, we have found the Half-Half Split,

ARL-Unbiased, and TPC (where ε1 = ε2 = ε and ε takes value of 0.1, 0.2,
0.3 or 0.5) designs and summarised the results in Table 2. For comparison, we
also plot the ARL curves in Figure 2 and Figure 3. These figures show that the
curves for ARL-Unbiased and TPC designs are close to each other, whereas they

Half Half Split

ARL Unbiased
400

TPC 0.1
TPC 0.2
TPC 0.3
TPC 0.5
300
ARL(rho)
200
100
0

0.0 0.5 1.0 1.5 2.0

rho
Figure 2. ARL curves for various designs when n = 5 and α = 0.0027. “TPC 0.i” denotes the TPC
design with ε1 = ε2 = 0.i for i = 1, 2, 3, 5.
312 LINGYUN ZHANG – K. GOVINDARAJU

are visibly different from the curve for Half-Half Split Design—the discrepancy
becomes less obvious when n = 10 than when n = 5. From our observations
we conclude that the ARL-Unbiased and TPC designs bring in more balance
between the increase and decrease in process standard deviation than the Half-
Half Split Design does. Hence we recommend the ARL-Unbiased design or
TPC design in practice when n is small (e.g. n = 4 or 5).
400

Half Half Split

ARL Unbiased
TPC 0.1
TPC 0.2
300

TPC 0.3
TPC 0.5
ARL(rho)
200
100
0

0.0 0.5 1.0 1.5 2.0

rho
Figure 3. ARL curves for various designs when n = 10 and α = 0.0027. “TPC 0.i” denotes the TPC
design with ε1 = ε2 = 0.i for i = 1, 2, 3, 5.

Consider the following finer points for choosing between the ARL-Unbiased
design and TPC design.
• They are both for correcting the inherent bias in ARL resulting from the
skew of the distribution of the control statistic for small subgroup sizes.
When subgroup size is large, say n ≥ 30, the skew becomes less severe
and we can simply use the S chart from the Half-Half Split Design. If a
decrease in process standard deviation is not of interest, an upper-sided S
chart (i.e. c1 = 0 and c2 = χ1−α,
2
n−1 ) is suitable.
• ARL-Unbiased design and TPC design are two natural but different methods
to assign control limits for S charts. If ε1 and ε2 are properly chosen,
 On probability limits for phase II S control chart 313

TPC design usually results in a quite ARL-Unbiased S chart (for this see
Figure 2 and Figure 3), although strictly speaking it is still ARL-biased.
• TPC design has the disadvantage that practitioners have to choose proper
values of ε1 and ε2 . However, TPC design is useful for achieving balanced
and small ARL values at specified values of ρ such as ρ = 0.5 and ρ = 1.5.

Table 2: Various designs for n = 5, 10 and α = 0.0027. The ε1 = ε2 = ε is for TPC designs.

ε n=5 n = 10

αl c1 c2 αl c1 c2

Half-Half Split − 0.00135 0.10577 17.80041 0.00135 1.24125 27.09313

ARL-Unbiased − 0.00222 0.13646 20.10776 0.00199 1.36912 28.77819

TPC 0.1 0.00198 0.12873 19.20813 0.00185 1.34416 28.31079

0.2 0.00209 0.13211 19.55236 0.00197 1.36533 28.70000
0.3 0.00218 0.13518 19.92702 0.00202 1.37355 28.87355
0.5 0.00223 0.13666 20.13753 0.00178 1.33139 28.10786

5. Concluding remarks

The reason why 3-sigma control limits should not be used for Phase II
S charts stems out from the skew nature of the distribution of S for small
subgroup sizes. Control limits must be obtained as probability limits. If the
subgroup size n is small (say not larger than 10), we recommend the use
of ARL-Unbiased design or TPC design that are discussed in Sections 3.2
and 3.3. R programs for ARL-Unbiased design and TPC design are available
in Appendix D.
The suggested probability limit designs require that the true process stan-
dard deviation σ0 is estimated correctly during the Phase I study. For a dis-
cussion of the size of observations to make a good estimate of σ0 in Phase I
study, we refer readers to Chen (1998) and Zhang et al. (2005).

Appendix A: Derivation of (3)

First, let us recall our notation. The process standard deviation is assumed
to
x be σ = ρσ0 ; s = (n − 1)/2, where n is the subgroup size; and F(x) =
0 (1/ (s)) u s−1 e−u du.
The run length of an S chart follows a geometric distribution and its mean
is equal to the reciprocal of the tail probability. We need to obtain the tail
314 LINGYUN ZHANG – K. GOVINDARAJU

probability, 1 − Pr(aσ0 < Si < bσ0 ). Under the assumptions of independence

and normality, (n − 1)Si2 /σ 2 has a chi-square distribution with n − 1 degrees of
freedom, which is the same as the gamma distribution, Gamma((n − 1)/2, 2) =
Gamma(s, 2). Noting that

(n − 1)Si2 (n − 1)Si2 σ 2 (n − 1)Si2 2

= = ρ ∼ Gamma(s, 2ρ 2 ) ,
σ02 σ2 σ02 σ2

we have
 (n − 1)Si2 
Pr(aσ0 < Si < bσ0 ) = Pr (n − 1)a 2 < < (n − 1)b2
σ0
2

 (n − 1)Si2 2 
= Pr (n − 1)a 2 < ρ < (n − 1)b 2
σ2
 (n − 1)a 2
(n − 1)Si2 (n − 1)b2 
= Pr < <
2ρ 2 2σ 2 2ρ 2
= F(sb /ρ ) − F(sa /ρ ) .
2 2 2 2

Thus,
1
ARL(ρ) = # $.
1 − F(sb2 /ρ 2 ) − F(sa 2 /ρ 2 )

Appendix B: Derivation of (5)

The first equation in (5) is the requirement of criterion (C1). The following
is an explanation for the second equation in  (5).
In Section 1 we defined that F(x) = 0x (1/ (s)) u s−1 e−u du. By (3), the
ARL function is
1
ARL(ρ) = # 
c2
 
c1
$ ,
1− F 2ρ 2
−F 2ρ 2

which has the same monotonicity as the function g(ρ) = F( 2ρc22 )− F( 2ρc12 ) does,
and which attains its maximum at where g(ρ) does. The derivative of g(ρ) is





  c2 c2 c1 c1
g (ρ) = F − 3 − F − 3
2ρ 2 ρ 2ρ 2 ρ

c1 c2 
1 s − 2ρ 2 s − 2ρ 2
= s−1 2s+1 ce − c2 e .
2 ρ (s) 1
 On probability limits for phase II S control chart 315

Setting g  (ρ) = 0 we get the solution

,
- c1 − c2
-∗
ρ =.  .
2s ln(c1 ) − ln(c2 )
Noting the fact that 0 < c1 < c2 , we can easily verify that
 > 0, if ρ < ρ ∗ ,

g  (ρ) = 0, if ρ = ρ ∗ ,

< 0, if ρ > ρ ∗ .
This means that g(ρ) firstly monotonically increases, attains its maximum at ρ ∗ ,
and after that it monotonically decreases. Thus, letting ρ ∗ be 1 or equivalently
c1 − c2
  =1
2s ln(c1 ) − ln(c2 )
is required for an ARL-Unbiased design.

Appendix C: Proof of the Theorem

Let F −1 (·) denote the inverse function of F(·) that is defined in Section
1. Because c1 = χα2l , n−1 and c2 = χ1−α
2
r , n−1
, we have

c1 = 2F −1 (αl ) ,
c2 = 2F −1 (1 − αr ) = 2F −1 (1 − α + αl ) .
We define
 
F −1 (1 − α + αl ) F −1 (αl )
p(αl ) = F −F , for 0 < αl < α .
ρ2 ρ2
Note that 1− p(αl ) is the tail probability, and its reciprocal is equal to the ARL.
To prove the theorem, it is equivalent to show that: p(αl ) is monotonically
increasing as αl increases, if ρ > 1; and p(αl ) is monotonically decreasing as
αl increases, if 0 < ρ < 1. Again, we resort to differentiation. Denote the
derivative of F(·) by f (·), and then the derivative of p(αl ) is

 F −1 (1 − α + αl ) 1 1
p (αl ) = f  
ρ2 ρ f F (1 − α + αl )
2 −1

F −1 (αl ) 1 1
− f  
ρ2 ρ f F (αl )
2 −1



F −1 (1−α+αl )   F −1 (αl )  
f ρ2
f F −1 (αl ) − f ρ2
f F −1 (1 − α + αl )
1
=     .
ρ2 f F −1 (αl ) f F −1 (1 − α + αl )
316 LINGYUN ZHANG – K. GOVINDARAJU

Because f (x) = x s−1 e−x / (s), the numerator in p  (αl ) becomes

 s−1
F −1 (αl )F −1 (1 − α + αl )  
e−I1 − e−I2 ,
((s))2 ρ 2(s−1)
where
F −1 (1 − α + αl )
I1 = F −1 (αl ) + ,
ρ2
and
F −1 (αl )
I2 = + F −1 (1 − α + αl ) .
ρ2
Now we see that the sign (positive or negative) of p  (αl ) is determined by
−I1 − (−I2 ) = I2 − I1
# $
(ρ 2 − 1) F −1 (1 − α + αl ) − F −1 (αl )
= .
ρ2

Since F −1 (1 − α + αl ) − F −1 (αl ) > 0, the sign of I2 − I1 is determined by

ρ 2 − 1. At this point we complete the proof.

Appendix D: R Programs

R Program for ARL-Unbiased Design

arL.unbiased.design<-function(n=5, my.alpha=0.0027) { my.s<- ( n - 1
) / 2

obj.func<-function(k)
{y<- pgamma(q=my.s*log(k)/(k-1), shape=my.s) -
pgamma(q=my.s*k*log(k)/(k-1), shape=my.s) - (1 - my.alpha)
y
}

# It can be shown that obj.func is a decreasing function

# We use a bi-section algorithm to find solution to obj.func(k)=0
p1<- 0.0
p2<- 1.0

while(1)
{midp<- (p1 + p2)/2
 On probability limits for phase II S control chart 317

result<- obj.func(k=midp)
if(abs(result) <= 1e-8) break
else
{ if (result > 1e-8) p1<-midp
else p2<-midp
}
}

my.c2<- 2 * my.s * log(midp)/( midp - 1.0 )

my.c1<- midp * my.c2

my.alpha1<- pgamma(q=my.c1/2.0, shape=my.s)

list( round( c(my.alpha1, my.c1, my.c2), 5) )

}

R Program for TPC Design

tpc.design<-function(n=5, my.alpha=0.0027) { # specify epsilon 1 and
epsilon 2
my.eps1<- 0.5
my.eps2<- 0.5

my.s<-(n-1)/2

up.bound<- 2*qgamma(p=my.alpha, shape=my.s)

my.obj.fun<-function(x)
{ my.c1<- x
my.c2<- 2*qgamma(p=1-my.alpha+pgamma(my.c1/2,shape=my.s),shape=my.s)
my.re<- 1.0/( 1 - ( pgamma(my.c2/2.0/(1-my.eps1)**2, shape=my.s) -
pgamma(my.c1/2.0/(1-my.eps1)**2,shape=my.s))) +
1.0/( 1 - ( pgamma(my.c2/2.0/(1+my.eps2)**2, shape=my.s) -
pgamma(my.c1/2.0/(1+my.eps2)**2, shape=my.s)))
my.re
}

my.c1<-optimize(f = my.obj.fun, interval=c(0, up.bound)) $minimum

my.alpha1<- pgamma(q=my.c1/2.0, shape=my.s)

my.alpha2<- my.alpha - my.alpha1

my.c2<- 2*qgamma(p=1-my.alpha2, shape=my.s)

318 LINGYUN ZHANG – K. GOVINDARAJU

# values of alpha l, c1 and c2 are returned

list( round( c(my.alpha1, my.c1, my.c2), 5) )
}

Acknowledgments

We thank the two anonymous referees for their critical comments and useful suggestions.

REFERENCES

Acosta-Mejia, C. A. (1998) Monitoring Reduction in Variability with the Range, IIE Transactions,
30, 515–523.
Champ, C. W. and Lowry, C. A. (1994) Adjusting the S-Chart for Detecting Both Increases and
Decreases in the Standard Deviation, Proceedings of the Decision Sciences Institute Annual
Conference, Honolulu, Hawaii, November 20-22, Vol. 3, 2112–2114.
Chen, G. (1998) The Run Length Distributions of the R, s and s 2 Control Charts when σ is Estimated,
The Canadian Journal of Statistics, 26, 311–322.
Chen, G., Cheng, S. W. and Xie, H. (2001) Monitoring Process Mean and Variability With One
New EWMA Chart, Journal of Quality Technology, 33, 223–233.
Devore, J. and Farnum, N. (2005) Applied Statistics for Engineers and Scientists, 2nd ed., Curt
Hinrichs.
Montgomery, D. C. (2005) Introduction to Statistical Quality Control, 5th ed., John Wiley & Sons,
Inc.
Pignatiello J. J. Jr., Acosta-Mejia C. A. and Rao, B. V. (1995) The Performance of Control
Charts for Monitoring Process Dispersion, Proceedings of the 4th Industrial Engineering
Research Conference, 320–328.
R development core team (2006) R: A language and environment for statistical computing. R Foun-
dation for Statistical Computing, Vienna, Austria. ISBN 3-900051-07-0, URL http://www.R-
project.org.
Ryan, T. P. (2000) Statistical Methods for Quality Improvement, 2nd ed., Wiley, New York, NY.
Woodall, W. H. and Montgomery D. C. (1999) Research Issues and Ideas in Satistical Process
Control, Journal of Quality Technology, 31, 376–387.
Zhang, L., Govindaraju, K., Bebbington, M. and Lai, C. D. (2004) On the Statistical Design of
Geometric Control Charts, Quality Technology and Quantitative Management, 1 (2), 233–
243.
Zhang, L., Bebbington, M. S., Lai, C. D. and Govindaraju, K. (2005) On Statistical Design of
the S 2 Control Chart, Communications in Statistics–Theory and Methods, 34, 229–244.

LINGYUN ZHANG K. GOVINDARAJU

IIST, Massey University IIST, Massey University
Private Box 756 Private Bag 11222
Wellington, New Zealand Palmerston North, New Zealand
[email protected] [email protected]