Filomat 29:10 (2015), 2325–2338 Published by Faculty of Sciences and Mathematics,

DOI 10.2298/FIL1510325V University of Niš, Serbia

Available at: http://www.pmf.ni.ac.rs/filomat

Economic Statistical Design of X Bar Control Chart for Non-Normal

Symmetric Distribution of Quality Characteristic

Kristina Veljkovica , Halima Elfaghihea , Vesna Jevremovica

a Department of Probability and Statistics, Faculty of Mathematics, University of Belgrade, Studentski trg 16, Belgrade, Serbia

Abstract. In economic statistical design of a control chart, the economic-loss function is minimized subject
to a constrained minimum value of power, maximum value of probability of false alarms and average
time to signal an expected shift. This paper is concerned with the optimum economic statistical design
of the X bar chart when quality characteristic has non-normal symmetric distribution. We considered
three types of distributions: Student distribution, standard Laplace distribution and logistic distribution.
For each of these distributions, we calculated theoretical distribution of standardized sample mean (or its
best approximation) and approximated it with normal, Pearson VII and Johnson SU distributions. For
considered example, constrained minimization of expected loss function was done using genetic algorithm
in statistical software R. We compared results of economic statistical design of X-bar chart for theoretical
distribution of standardized sample mean with the results for normal, Pearson and Johnson distributions.
We found that, for all chosen distributions of quality characteristic, Pearson VII distribution and Johnson
SU distribution give results very close to results based on theoretical distribution of standardized sample
mean, while normal distribution gives much worse fit.

1. Introduction

The X bar chart is extensively used in practice to monitor a change in the process mean. It is usually
assumed that quality characteristic under the surveillance of an X-chart has a normal distribution. On the
other hand, occurrence of non-normal data in industry is quite common (see Alloway and Raghavachari,
1991; Janacek and Meikle, 1997). Violation of normality assumption results in incorrect control limits of
control charts (Alwan, 1995). Misplaced control limits lead to inappropriate charts that will either fail to
detect real changes in the process or which will generate spurious warnings when the process has not
changed.
In the case of non-normal symmetric distribution of quality characteristics, no recommendations, except
the use of normal distribution, are given in the quality control literature. Approximation of the distribution
of sample mean with normal distribution is based on the central limit theorem, but in practice small sample
sizes are usually used.
We will consider three types of non-normal symmetric distributions of quality characteristic: Student
distribution, Laplace distribution and logistic distribution. These distributions are chosen because of

2010 Mathematics Subject Classification. 60E05; 62E17; 62P30; 90C59

Keywords. X-bar control chart, economic statistical design, non-normal symmetric distribution, Pearson type VII distribution,
Johnson SU distribution
Received: 01 September 2014; Accepted: 07 January 2015
Communicated by Dragana Cvetković-Ilić
Email address: [email protected] (Kristina Veljkovic)
 K.Veljkovic et al. / Filomat 29:10 (2015), 2325–2338 2326

their applications in various disciplines (economics, finance, engineering, hydrology, etc, see for instance
Ahsanullah et al., 2014; Balakrishnan, 1991; Kotz et al., 2001). For each of these distributions, we calculated
theoretical distribution of standardized sample mean (or its best approximation) and then we approximated
it with normal, Pearson VII and Johnson SU distributions. Pearson system of distributions, Johnson’s
transformations and normal distribution are known to provide approximations to a wide variety of observed
distributions (Johnson et al., 1994).
In this paper, we will be concerned with the optimum economic statistical design of the X-bar chart for
chosen distributions of quality characteristic. In economic statistical design of a control chart, the economic-
loss function is minimized subject to a constrained minimum value of power and maximum value of the
Type I error probability and ATS (average time to signal) an expected shift (Saniga, 1989). Constrained
minimization of expected loss function was done in statistical software R, using genetic algorithm ga from R
package GA (Scrucca, 2013). Results of economic statistical design of X-bar chart for theoretical distribution
of standardized sample mean and normal, Pearson VII and Johnson SU approximations are compared, with
results based on the theoretical distribution of sample mean as the ’golden’ standard.
Our objective is, using previous theoretical results concerning distributions of standardized sample
mean, to provide a reliable method for design of X bar control chart that takes into account economic factors
while achieving desirable statistical properties in the case when distribution of the data is symmetric, but
non-normal.
The rest of this paper is organized as follows: basic assumptions about the process and methods of
design of X-bar control chart are discussed in Section 2. Description of distribution of quality characteristic
and distribution of standardized sample mean is given in Sections 3 and 4, respectively. Normal, Pearson
and Johnson approximations are considered in Section 5. An illustrative example on the construction of
the economic statistical design of X bar chart for theoretical distribution of standardized sample mean and
its normal, Pearson VII and Johnson SU approximations is given in Section 6. Conclusions are drawn in
Section 7.

2. Economic statistical design of the X bar chart

2.1. Basic assumptions

We shall state briefly the basic assumptions about the process. It is presumed that a process begins in
in-control state with mean µ0 and that single assignable cause of magnitude δ results in a shift in the process
mean from µ0 to either µ0 − δσ or µ0 + δσ, where σ is a standard deviation which is assumed to remain
stable.
The process is monitored
√ by the X√bar chart with central line µ0 and upper and lower control limits,
respectively, µ0 + kσ/ n and µ0 − kσ/ n, where n represents the sample size and k width of control limits.
Samples are taken from the process every h hours and the sample mean is plotted on the X-bar chart. If a
sample mean exceeds control limits, it is assumed that some shift in the process mean has occurred and a
search for the assignable cause is initiated. The process is allowed to continue operating during the search
for the assignable cause. The parameters µ0 , δ and σ are assumed to be known.
The assignable cause is assumed to occur according to a Poisson process with an intensity of λ occur-
rences per hour. That is, assuming that the process begins in the in-control state, the time interval that
the process remains in control is an exponential random variable with mean λ1 h. Therefore, given the
occurrence of the assignable cause in the interval between the jth and (j + 1)st sample, the expected time of
occurrence within the interval between samples is
R ( j+1)h
jh
(t − jh)λe−λt dt 1 − (1 + λh)e−λh
τ= = .
λ(1 − e−λh )
R ( j+1)h
jh
λe−λt dt
 K.Veljkovic et al. / Filomat 29:10 (2015), 2325–2338 2327

2.2. Design parameters

Various kinds of parameters are considered in the economic statistical design of the control chart and
they are denoted by following symbols:

• n: Sample size.

• h: Sampling interval.

• k: Width of control limits.

• σ: Standard deviation of observations.

• µ0 : In-control process mean.

• µ1 : Out-of control process mean.

• δ: Shifts in process mean in standard deviation units when assignable cause occurs (δ = |µ1 − µ0 |/σ).

• λ: It is assumed that in-control time follows an exponential distribution with mean λ1 .

• τ: Expected time of occurrence of the assignable cause, given it occurs between the j-th and ( j + 1)-st
samples.

• α: Probability of type I error.

• β: Probability of type II error.

• ATS: Average time to signal an expected shift, ATS = h

1−β .

• 1: The time required to take a sample and interpret the results.

• D: The time required to find an assignable cause following an out-of-control signal.

• a1 : Fixed cost per sample.

• a2 : Variable cost per unit sampled.

• a3 : Cost for searching and repairing the assignable cause.

0
• a3 : Cost of investigating a false alarm.

• V0 : Profit per hour earned by the process operating in-control.

• V1 : Profit per hour earned by the process operating out-of-control

(V0 > V1 ).

• a4 : Hourly penalty cost associated with production in the out-of-control state, a4 = V0 − V1

−λh
• κ: Expected number of false alarms, κ = α 1−e
e
−λh .

• E(T): Expected length of a production cycle. The production cycle time T is defined as the interval of
time from the start of production (the process is assumed to start in the in-control state), following an
adjustment to the detection and elimination of the assignable cause.

• E(C): Expected cost per cycle. Total cost C consists of the costs of testing and sampling, the costs asso-
ciated with investigating an out-of-control signal and with the repair or correction of any assignable
causes found and the costs associated with the production of nonconforming items.
E(T)
• s: Expected number of samples taken within a cycle, s = h .
 K.Veljkovic et al. / Filomat 29:10 (2015), 2325–2338 2328

2.3. The expected cycle time

Each production cycle begins with the process in the in-control state and continues until process moni-
toring via control charts results in an out-of-control signal. Following an adjustment in which the process
is returned to the in-control state, a new cycle begins.
Length of a production cycle T is partitioned into five time intervals, Ia , Ib , Ic , Id , and Ie (Figure 1).

Figure 1: Intervals of production cycle

• Ia : The time until the assignable cause occurs from the start of the cycle.

• Ib : The time until the next sample is taken from the end of Ia .

• Ic : The time until the chart gives an out-of-control signal from the end of Ib .

• Id : The time to analyze the sample and chart the result from the end of Ic .

• Ie : The time to discover the assignable cause and repair the process from the end of Id .

Three time points are denoted on Figure 1: point 1 represents last sample before the occurrence of the
assignable cause, point 2 first sample taken after assignable cause and point 3 the moment when lack of
control is detected. Process remains in-control state during time interval Ia and it is in out-control state
during intervals Ib , Ic , Id and Ie .
Expected lengths of each of these time intervals are, respectively, equal to

1
E(Ia ) = , E(Ib ) = h − τ, E(Ic ) = ATS − h, E(Id ) = 1n, E(Ie ) = D.
λ

Then, expected length of a cycle is

1 h
E(T) = + − τ + 1n + D.
λ 1−β

2.4. The expected cost

Expected cost during a production cycle is equal to
!
1 h 0
E(C) = V0 · + V1 · − τ + 1n + D − a3 − a3 κ − (a1 + a2 n)s
λ 1−β
 K.Veljkovic et al. / Filomat 29:10 (2015), 2325–2338 2329

2.5. The expected loss per hour

The expected loss function is equal to
  0
E(C) a1 + a2 a4 1−β − τ + 1n + D + a3 + a3 κ
h

E(L) = V0 − = + .
E(T) h 1
+ h − τ + 1n + D
λ 1−β

2.6. Methods of design

The use of control chart requires one to select a sample size, a sampling frequency and control limits for
the chart. Selection of these three parameters is usually called the design of the control chart. Currently,
there are three general methods for design of control charts: statistical, economic and economic statistical
design. In statistically designed control charts, control-limit width (which determines the Type I error
probability) and power are preselected. Then, sample size and sampling frequency (if ATS is specified) can
be determined (Woodall, 1985).
The objective of economic design is to find the sample size, control-limit width, and sampling frequency
that minimize the loss in profit composed of several given costs. Duncan (1956) first proposed an economic
model for the optimum economical design of the X-bar control chart. This research stimulated many
subsequent studies on the subject. Lorenzen and Vance (1986) proposed a general method for determining
the economic design parameters of control charts.
The weakness of statistically designed control charts are high operating costs. On the other hand,
economic design does not consider the probability of type I and type II errors which can lead to high
probability of false alarms and low power. Saniga (1989) proposed economic statistical model which
combines properties of economic and statistical design: it is a method for economic design of control charts
that have bounds on type I and type II error probabilities and the average time to signal (ATS) an expected
shift. Economic statistical design considers economic factors while achieving desirable statistical properties
and, therefore, represents the improvement of economic or statistical design.

2.7. Economic statistical model

The economic statistical design of a control chart is defined as a design in which the expected economic-
loss cost function E(L) is minimized subject to a constrained minimum value of power and maximum value
of the Type I error probability and ATS an expected shift. We wish to find sample size, control-limit width,
and sampling frequency to

minimize E(L) (1)

subject to
α ≤ α0 , p ≥ p0 , ATS ≤ ATS0 ,

where p = 1 − β, α0 , p0 and ATS0 are, respectively, power and the desired bounds on the Type I error
probability, power, and ATS an expected shift.

3. Distribution of quality characteristic

In this paper we consider three types of non-normal symmetric distributions of quality characteristic X:
Student’s distribution, Laplace (double exponential) and logistic distribution (Ðorić et al. 2007, Johnson et
al. 1994, Johnson et al. 1995). All chosen distributions have heavier tails than normal distribution.
Distributions are given by their density functions. We will use following notations for mean, variance,
E(X−E(X))3
skewness and kurtosis of quality characteristic, respectively: µ = E(X), σ2 = Var(X), α3 = 3 ,
σ2
4
E(X−E(X))
α4 = σ4
. As all chosen distributions are symmetric around the mean, α3 = 0.
 K.Veljkovic et al. / Filomat 29:10 (2015), 2325–2338 2330

1. Student’s distribution t(ν, µ, η)

!2 − ν+1
Γ ν+1
  
1 x − µ 
2
2
f (x) = √   1 + , x ∈ R,


η πνΓ 2ν ν η


where ν (ν ∈ R) is the shape parameter, η (η > 0) is scale parameter and µ (µ ∈ R) location parameter.
Variance and kurtosis are, respectively, equal to

ν 2 6
σ2 = η , ν > 2, α4 = 3 + , ν > 4.
ν−2 ν−4
X−µ
Random variable Y = η has Student distribution t(ν) with density function

!− ν+1
Γ ν+1
 
2 y2 2
f (y) = √   1+ , y ∈ R.
πνΓ 2ν ν

2. Laplace distribution L2 (η, µ)

|x − µ|
( )
1
f (x) = exp − , x ∈ R,
2η η

where µ (µ ∈ R) is location parameter and η (η > 0) scale parameter. Variance and kurtosis are,
respectively, equal to

σ2 = 2η2 , α4 = 6.

Random variable Y = X − µ has standard Laplace distribution L1 (η) with density function
( )
1 |y|
fY (y) = exp − , y ∈ R.
2η η

3. Logistic distribution LGS(µ, η)

n x−µ o
exp − η
f (x) =  n x−µ o2 , x ∈ R,
η · 1 + exp − η

where µ (µ ∈ R) is the location parameter and η (η > 0) scale parameter. Variance and kurtosis are,
respectively, equal to

η2 π2
σ2 = , α4 = 4.2.
3
X−µ
Random variable Y = η has standard logistic distribution with density function

e−y
fY (y) = , y ∈ R,
(1 + e−y )2
 K.Veljkovic et al. / Filomat 29:10 (2015), 2325–2338 2331

4. Distribution of standardized sample mean

4.1. Sample from Student’s distribution
Witkowský (2001, 2004) proposed a method for numerical evaluation of the distribution function and
the density function of a linear combination of independent Student’s t random variables. The method is
based on the inversion formula which leads to the one-dimensional numerical integration.
X −µ
Let (X1 , X2 , . . . , Xn ) be a sample from Student’s t(ν, µ, λ) distribution. Then Yk = kη , k = 1, 2, . . . , n are
independent Student’s tν random variables with ν degrees of freedom. Further, let Y = nk=1 Yk be sum of
P
these variables and let φYk (t) denote the characteristic function of Yk . The characteristic function of Y is
n
Y
φY (t) = φYk (t),
k=1

where
1  1  2ν  1 
φYk (t) = ν ν
ν 2 |t| Kν/2 ν 2 |t| ,
2 2 −1Γ( 2 )
where Kα (z) denotes modified Bessel function of the second kind.
For more details see Witkovský (2001). Note that the characteristic function of the Student’s random
variable is a real function.
The cumulative distribution function (cdf) FY (y) of random variable Y is according to the inversion
formula due to Gil-Pelaez (1951) given by

e−ity φY (t)
∞
Z !
1 1
FY (y) = − Im dt =
2 π 0 t
Z ∞
1 1 sin (ty)φY (t)
= + dt (2)
2 π 0 t
and the probability density function (pdf) fY (y) of Y is given by
Z ∞
1
f (y) = Re(e−ity φY (t)) dt =
π 0
Z ∞
1
= cos (ty)φY (t) dt (3)
π 0
For any chosen y algorithm tdist in R package tdist (Witkovsky and Savin, 2005) evaluates the integrals in
(2) and (3) by multiple p-points Gaussian quadrature over the real interval t ∈ (0, 10π). The whole interval is
divided into m subintervals given by pre-specified limits and the integration over each subinterval is done
with p-points Gaussian quadrature which involves base points bi j , and weight factors wi j , i = 1, 2, . . . , p, j =
1, 2, . . . , m. So,
 
m p
1 1 X X sin bi j y
FY (y) ≈ + wi j φY (bi j ) =
2 π bi j
j=1 i=1
 
m X p
1 1 X sin b i j y
= + Wi j ,
2 π bi j
j=1 i=1
p
m X
1 X  
fY (y) ≈ cos bi j y wi j φY (bi j ) =
π
j=1 i=1
m p
1 XX  
= cos bi j y Wi j ,
π
j=1 i=1
 K.Veljkovic et al. / Filomat 29:10 (2015), 2325–2338 2332

where Wij = wi j φY (bi j ). For evaluation of FY (y) and fY (y) in many different points the algorithm requires
only one evaluation of the weights Wi j , for i = 1, 2, . . . , p and j = 1, 2, . . . , m, which directly depend on the
characteristic function φY (·) and do not depend on y.
X−µ √
q
Then, distribution function and density function of standardized sample mean Tn = σ n = Y ν−2
nν
are equal to
 r 
ν−2
r !
  nν
FTn (t) = P  = FY
 
Y ≤ t t
 

 nν 
 ν−2
r r !
nν nν
fTn (t) = · fY t , t∈R
ν−2 ν−2
Skewness and kurtosis of standardized sample mean are, respectively, equal

6
α3,Tn = 0, α4,Tn = 3 + .
n(ν − 4)
Probability of I type error and power are, respectively, equal to
√ √
α = 2FTn (−k), 1 − β = FTn (−k − δ n) + FTn (−k + δ n).

4.2. Sample from Laplace distribution

Let (X1 , X2 , . . . , Xn ) be a sample from Laplace distribution L2 (µ, η). Then, Yk = Xk − µ, k = 1, 2, . . . , n are
independent random variables with standard Laplace distribution L1 (η).
We will use the property that if two independent random variables have exponential ε(η) distribution,
then their difference has standard Laplace distribution. Further, standard exponential distribution is gamma
distribution, Γ(1, η). Sum of n independent variables with Γ(1, η) distribution is gamma distribution Γ(n, η).
In that way, we conclude that sum of n independent random variables Y1 , Y2 , . . . , Yn with standard Laplace
distribution can be written as the difference of two random variables with gamma distribution Γ(n, η) which
is called bilateral gamma distribution.
Bilateral gamma distribution is symmetric around 0, with density function for y > 0 (Küchler and Tappe,
2008)
 η n n−1
1 X
f (y) = · ak yk e−ηy ,
2 (n − 1)!
k=0

where the coefficients (ak )k=0,...,n−1 are given by

! n−2−k
n−1 1 Y
ak = n−1−k
(n + l), an−1 = 1.
k (2η)
l=0

Probability distribution function, for y > 0 is

 η n n
1 1 X ak
FY (y) = + · γ(k + 1, ηx),
2 2 (n − 1)! ηk+1 k=0

where γ(n, x) is lower incomplete gamma function.

For evaluation of cdf we will use following recursive relation for incomplete gamma function

γ(n, x) = (n − 1)γ(n − 1, x) − xn−1 e−x

and γ(1, x) = 1 − e−x .

 K.Veljkovic et al. / Filomat 29:10 (2015), 2325–2338 2333
X−µ √
Then, distribution function and density function of standardized sample mean Tn = σ n= Y
√
η· 2n
 
 Y
   √ 
FTn (t) = P  = FY η 2nt ,
 
√ ≤ t

 η · 2n
 

√  √ 
fTn (t) = η 2n · fY η 2nt , t ∈ R
Skewness and kurtosis of standardized sample mean are, respectively, equal
3
α3,Tn = 0, α4,Tn = 3 + .
n
Probability of I type error and power are, respectively equal to
√ √
α = 2FTn (−k), 1 − β = FTn (−k − δ n) + FTn (−k + δ n).

4.3. Sample from logistic distribution

Let (X1 , X2 , . . . , Xn ) be a random sample from logistic distribution LGS(µ, η).
George and Mudholkar (1983) showed that a standardized Student’s t-distribution provides a very
good approximation for the distribution of a convolution of n i.i.d. logistic variables. These authors then
compared three approximations: (1) standard normal approximation, (2) Edgeworth series approximation
correct to order n−1 , and (3) Student’s t approximation with ν = 5n + 4 degrees of freedom. They showed
that the third provides a very good approximation.
Gupta and Han (1992) considered the Edgeworth series expansions up to order n−3 for the distribution
of the standardized sample mean

X−µ√
Tn = n.
σ
Distribution function is given by
1 1 6 1 1 48
   
FTn (t) ≈ Φ(t) − ϕ(t) · H3 (t) + 2 · H5 (t)+
n 4! 5 n 6! 7
35 6 2
!!
1 1 432 210 48 6
  
+ H7 (t) + 3 · H7 (t) + · H9 (t)+
8! 5 n 8! 5 10! 7 5
5775 6 3
  !!
+ H11 (t) , t ∈ R,
12! 5
where ϕ(·) and Φ(·) are standard normal pdf and cdf and H j (x) is the Hermite polynomial.
Then, probability density function of standardized sample mean is

1 1 6 
 
0
fTn (t) ≈ ϕ(t) 1 + · · · tH3 (t) − H3 (t) +
n 4! 5
 35  6 2 
!
1 1 48  0 0

+ · · tH5 (t) − H5 (t) + tH7 (t) − H7 (t) +
n2 6! 7 8! 5
1 1 432   210 48 6 
 
0 0
+ · · tH 7 (t) − H 7 (t) + · tH 9 (t) − H 9 (t) +
n3 8! 5 10! 7 5
5775 6 3 
  !!
0

+ tH11 (t) − H11 (t) , t ∈ R
12! 5
Gupta and Han (1992) compared this approximation with the approximations mentioned earlier, and
they showed the approximation to be far better than even the Student’s t-approximation suggested by
George and Mudholkar (1983).
 K.Veljkovic et al. / Filomat 29:10 (2015), 2325–2338 2334

Skewness and kurtosis of standardized sample mean are, respectively, equal

1.2
α3,Tn = 0, α4,Tn = 3 + .
n
Probability of I type error and power are, respectively equal to
√ √
α = 2FTn (−k), 1 − β = FTn (−k − δ n) + FTn (−k + δ n).

5. Approximations of the distribution of sample mean

We will approximate distribution of standardized sample mean Tn with normal distribution, Pearson’s
type VII distribution and Johnson’s SU distribution.

5.1. Normal distribution

Probability of I type error and power are, respectively equal to
√ √
α = 2Φ(−k), 1 − β = Φ(−k − δ n) + Φ(−k + δ n),

where Φ(·) is a probability distribution function of standard normal distribution.

5.2. Pearson type VII distribution

Probability density function is equal to (Johnson et al. 1994, Ðorić et al. 2007)
!−m
1 x2
f (x) =  · 1+ 2 , x ∈ R,
α

αB m − 12 , 12
q q
5α4 −9 2α4 2α4
where m = 2(α4 −3)
, α = σT n α4 −3 = α4 −3 and B(a, b) is beta function.
Then, cumulative distribution function is equal to

1 1 1
 
F(x) = Iα2 /(α2 +x2 ) m − , ,
2 2 2
for x < 0 and
1 1 1
 
F(x) = 1 − Iα2 /(α2 +x2 ) m − , ,
2 2 2
B (a,b)
for x > 0, where Ix (a, b) = B(a,b)
x
and Bx (a, b) is incomplete beta function.
Probability of I type error and power are, respectively equal to
√ √
α = 2F(−k), 1 − β = F(−k − δ n) + F(−k + δ n).

5.3. Johnson SU distribution

In our case, Johnson SU distribution is defined by (Johnson 1949, Johnson et al. 1994)

Tn − ξ
!
Z = γ + ζ arcsinh ,
ψ

where Tn is standardized sample mean (with unbounded range) and arcsinh(·) is inverse sine hyperbolic
function. Transformed variable Z has standard normal distribution. Symbols γ, ζ, ξ and ψ represent
parameters of the distribution.
 K.Veljkovic et al. / Filomat 29:10 (2015), 2325–2338 2335

Using method of moments (see Hill, 1976), we get following estimates for parameters γ, ζ, ξ, ψ. We will
1
use notation ω = e ζ2 .
As α3,Tn = 0, the required fitted Johnson’s curve is symmetric, and
q

1 1
ω= 2α4,Tn − 2 2 − 1, ζ = (ln ω)− 2 , γ = 0.

Parameters ξ and ψ can be found from

s
2σ2Tn
r
2
ξ = µTn = 0, ψ = = .
ω2 − 1 ω2 − 1
Then, Johnson’s transformation is equal to
!
Tn
Z = ζ arcsinh .
ψ
Probability of I type error and power are, respectively equal to
!!
k
α = 2Φ −ζ arcsinh ,
ψ
√ !! √ !!
k+δ n −k + δ n
1 − β = Φ ζ arcsinh − Φ ζ arcsinh .
ψ ψ

6. Example
0
Let X represent measurable quality characteristic. Let a1 = $1, a2 = $0.1, a3 = $25, a3 = $50, a4 = $100,
λ = 0.05, δ = 2, 1 = 0.0167h, D = 1 (see Duncan,1956; Montgomery, 2005).
Constrained minimization of expected loss function (1) was done using genetic algorithm ga from R
package GA (Scrucca, 2013). We selected following values for bounds on type I error probability, power
and ATS an expected shift: α0 = 0.05, p0 = 0.9, ATS0 = 2h.
We considered four distributions of quality characteristics: Student distributions with 5 and 10 degrees
of freedom, logistic distribution and Laplace distribution with η = 1. Using results from Section 4, we
calculated theoretical distribution of standardized sample mean (or its best approximation) and, afterwards,
expressions for type I error probability, power and expected loss function. The same procedure was followed
using normal, Pearson VII and Johnson’s distributions. We considered samples of sizes 3 to 10, finding, by
constrained optimization of expected loss function, the optimal values of parameters k and h.
We provided to genetic algorithm ga interval of values for parameters k and h in the following way.
Solving equations α = α(k) = 0.05 and p = p(k) = 0.9 using Brent’s root-finding method (Brent, 1973), we
got values k1 and k2 . By the constraints (1) we got interval of values [k1 , k2 ] for parameter k and [0, 2p(k1 )]
for parameter h. Results of the minimization of expected loss function are given in Table 1.
By looking at the results in Table 1, we can make following conclusions.
1. Normal approximation of standardized sample mean gives smallest values for k,h, ATS and expected
loss and greatest power. One can not be automatically rely on these results, considering the fact that
theoretical distribution of standardized sample mean gives the most precise answer.
2. Student distribution with 10 degrees of freedom is very close to normal distribution, so it is expected
that results for normal approximation (optimal values of k and h, α, power, ATS and expected loss)
should be close to results based on theoretical distribution, which is confirmed. Normal approximation
does not fit so well in the case of Student distribution with ν = 5 degrees of freedom, logistic and
Laplace distribution.
3. Pearson’s VII and Johnson SU approximations give results which are very close to results based on
theoretical distribution of standardized sample mean, in all cases.
 2336

Distribution Theoretical distribution Normal distribution Pearson VII distribution Johnson SU distribution
k h α 1 − β ATS EL k h α 1 − β ATS EL k h α 1 − β ATS EL k h α 1 − β ATS EL
Student
ν=5
n=3 2.25 1.00 0.030 0.90 1.11 11.31 2.18 0.98 0.029 0.90 1.09 11.25 2.27 1.00 0.031 0.90 1.11 11.35 2.27 1.00 0.032 0.90 1.11 11.40
n=4 2.77 0.83 0.011 0.90 0.92 10.39 2.71 0.77 0.0067 0.90 0.86 10.13 2.78 0.82 0.012 0.90 0.91 10.46 2.78 0.84 0.013 0.90 0.93 10.50
n=5 3.17 0.73 0.0047 0.91 0.80 10.04 3.05 0.69 0.0023 0.92 0.75 9.84 3.21 0.72 0.0054 0.91 0.80 10.09 3.23 0.72 0.0054 0.90 0.80 10.11
n=6 3.35 0.72 0.0030 0.94 0.76 9.89 3.20 0.70 0.0014 0.96 0.74 9.74 3.40 0.72 0.0035 0.94 0.77 9.93 3.43 0.72 0.0034 0.93 0.77 9.94
n=7 3.52 0.72 0.0020 0.96 0.74 9.83 3.35 0.71 0.0082 0.97 0.73 9.71 3.59 0.72 0.0022 0.96 0.75 9.86 3.59 0.72 0.0022 0.96 0.75 9.86
n=8 3.68 0.72 0.0013 0.97 0.73 9.81 3.49 0.71 0.00049 0.99 0.72 9.73 3.74 0.71 0.0015 0.97 0.74 9.83 3.75 0.71 0.0014 0.97 0.74 9.83
n=9 3.83 0.72 0.00093 0.98 0.73 9.83 3.62 0.71 0.00030 0.99 0.72 9.77 3.90 0.72 0.00097 0.98 0.73 9.84 3.90 0.71 0.00095 0.98 0.73 9.84
n = 10 3.98 0.72 0.00063 0.99 0.72 9.87 3.75 0.71 0.00017 0.99 0.72 9.82 4.03 0.72 0.00066 0.99 0.73 9.87 4.03 0.71 0.00064 0.99 0.72 9.87
ν = 10
K.Veljkovic et al. / Filomat 29:10 (2015), 2325–2338

n=3 2.20 0.98 0.030 0.90 1.09 11.31 2.18 0.98 0.029 0.90 1.09 11.25 2.20 0.99 0.031 0.90 1.10 11.32 2.20 0.99 0.031 0.90 1.10 11.33
n=4 2.73 0.78 0.0084 0.90 0.86 10.23 2.71 0.74 0.0067 0.90 0.82 10.13 2.73 0.75 0.0085 0.90 0.83 10.23 2.73 0.77 0.0084 0.90 0.86 10.23
n=5 3.08 0.71 0.0032 0.92 0.77 9.90 3.05 0.69 0.0023 0.92 0.75 9.84 3.11 0.70 0.0030 0.92 0.76 9.90 3.10 0.70 0.0030 0.92 0.76 9.90
n=6 3.26 0.70 0.0018 0.95 0.74 9.78 3.20 0.70 0.0014 0.96 0.74 9.74 3.26 0.70 0.0018 0.95 0.74 9.78 3.26 0.70 0.0018 0.95 0.74 9.78
n=7 3.41 0.71 0.0011 0.97 0.73 9.74 3.34 0.71 0.00083 0.97 0.73 9.71 3.41 0.71 0.0011 0.97 0.73 9.74 3.41 0.71 0.0011 0.97 0.73 9.74
n=8 3.55 0.71 0.0067 0.98 0.72 9.75 3.49 0.71 0.00049 0.99 0.72 9.73 3.55 0.71 0.00066 0.98 0.72 9.75 3.55 0.71 0.00066 0.98 0.72 9.75
n=9 3.69 0.71 0.00041 0.99 0.72 9.78 3.62 0.71 0.00029 0.99 0.72 9.77 3.69 0.71 0.00041 0.99 0.72 9.78 3.69 0.71 0.00041 0.99 0.72 9.78
n = 10 3.82 0.71 0.00025 0.99 0.72 9.82 3.74 0.72 0.00018 0.99 0.72 9.82 3.82 0.71 0.00025 0.99 0.72 9.83 3.82 0.71 0.00025 0.99 0.72 9.83
Logistic
n=3 2.21 1.00 0.031 0.90 1.11 11.34 2.18 1.00 0.029 0.90 1.11 11.25 2.21 1.00 0.031 0.90 1.11 11.32 2.21 1.02 0.031 0.90 1.13 11.33
n=4 2.74 0.77 0.0088 0.90 0.85 10.26 2.72 0.74 0.0066 0.90 0.82 10.12 2.73 0.75 0.0089 0.90 0.83 10.26 2.72 0.77 0.0090 0.90 0.85 10.26
n=5 3.10 0.71 0.0032 0.92 0.77 9.92 3.04 0.69 0.0023 0.92 0.75 9.84 3.11 0.71 0.0032 0.92 0.77 9.92 3.08 0.70 0.0034 0.92 0.76 9.92
n=6 3.27 0.70 0.0019 0.95 0.74 9.79 3.20 0.70 0.0014 0.96 0.74 9.74 3.27 0.70 0.0019 0.95 0.74 9.79 3.27 0.71 0.0019 0.95 0.74 9.79
n=7 3.42 0.71 0.0011 0.97 0.73 9.75 3.35 0.71 0.00082 0.97 0.73 9.71 3.42 0.71 0.0011 0.97 0.73 9.75 3.42 0.71 0.0011 0.97 0.73 9.75
n=8 3.56 0.71 0.00069 0.98 0.72 9.75 3.49 0.71 0.00049 0.99 0.72 9.73 3.56 0.71 0.00070 0.98 0.72 9.75 3.56 0.71 0.00069 0.98 0.72 9.75
n=9 3.70 0.71 0.00042 0.99 0.72 9.78 3.62 0.71 0.00029 0.99 0.72 9.77 3.70 0.71 0.00043 0.99 0.72 9.78 3.70 0.71 0.00043 0.99 0.72 9.78
n = 10 3.83 0.71 0.00026 0.99 0.72 9.83 3.75 0.71 0.00018 0.99 0.72 9.82 3.83 0.71 0.00027 0.99 0.72 9.83 3.83 0.71 0.00027 0.99 0.72 9.83
Laplace
η=1
n=3 2.24 1.02 0.034 0.90 1.13 11.46 2.18 0.98 0.029 0.90 1.09 11.25 2.24 0.99 0.031 0.90 1.10 11.35 2.24 1.01 0.032 0.90 1.13 11.38
n=4 2.76 0.82 0.011 0.90 0.91 10.41 2.71 0.76 0.0066 0.90 0.84 10.13 2.75 0.79 0.011 0.90 0.87 10.37 2.76 0.78 0.011 0.90 0.87 10.38
n=5 3.18 0.70 0.0041 0.91 0.78 10.00 3.05 0.69 0.0023 0.92 0.75 9.84 3.16 0.71 0.0042 0.91 0.78 10.00 3.17 0.71 0.0042 0.91 0.79 10.00
n=6 3.36 0.71 0.0024 0.94 0.75 9.85 3.20 0.70 0.0014 0.96 0.73 9.74 3.34 0.71 0.0026 0.94 0.75 9.86 3.34 0.71 0.0025 0.94 0.76 9.86
n=7 3.41 0.71 0.0011 0.97 0.73 9.74 3.34 0.71 0.00083 0.97 0.73 9.71 3.41 0.71 0.0011 0.97 0.73 9.74 3.41 0.71 0.0011 0.97 0.73 9.74
n=8 3.55 0.71 0.00067 0.98 0.72 9.75 3.49 0.71 0.00049 0.99 0.72 9.73 3.55 0.71 0.00066 0.98 0.72 9.75 3.55 0.71 0.00066 0.98 0.72 9.75
n=9 3.69 0.71 0.00041 0.99 0.72 9.78 3.62 0.71 0.00029 0.99 0.72 9.77 3.69 0.71 0.00041 0.99 0.72 9.78 3.69 0.71 0.00041 0.99 0.72 9.78
n = 10 3.82 0.71 0.00025 0.99 0.72 9.82 3.74 0.72 0.00018 0.99 0.72 9.82 3.82 0.71 0.00025 0.99 0.72 9.83 3.82 0.71 0.00025 0.99 0.72 9.83
Table 1: Optimum solution to Example
 K.Veljkovic et al. / Filomat 29:10 (2015), 2325–2338 2337

Our results suggest use of reliable method of economic statistical design of X bar control chart based on
Pearson’s VII or Johnson’s SU approximations in the case when distribution of the data is symmetric but
non-normal. Survey of literature (see for instance Li and Lin, 2013; Wu and Govindaraju, 2014; Wu et al.,
2014) indicates possibilities of application of our method in statistical quality control.

7. Conclusions

We considered optimum economic statistical design of X bar control chart when quality characteristic
has one of the following non-normal symmetric distributions: Student distribution with 5 and 10 degrees of
freedom, logistic distribution and Laplace distribution with η = 1. We calculated theoretical distribution of
standardized sample mean (or its best approximation) and approximated it with normal, Pearson type VII
and Johnson SU distribution. Constrained minimization of expected loss function was done using genetic
algorithm ga from R package GA. For considered example, we got table of results for optimal values of
parameters k and h, values of type I error probability, power, ATS and minimal value of expected loss, for
sample sizes from 3 to 10. For all chosen distributions of quality characteristic, Pearson type VII distribution
and Johnson SU distribution provide results which are very close to results based on theoretical distribution
of standardized sample mean, while normal distribution gives much worse fit.
Although our results of economic statistical design of X-bar chart are based only on three types of non-
normal symmetric distributions, we would recommend use of Pearson type VII distribution or Johnson SU
distribution in general case, when distribution of quality characteristic is non-normal but symmetric. Other
cases will be in scope of our future research.

References
[1] Ahsanullah, M., Golam Kibria B.M. and Shakil, M.(2014). Normal and Student’s t Distributions and Their Applications, Atlantis Press,
Paris.
[2] Balakrishnan, N. (ed.) (1992). Handbook of the Logistic Distribution, Dekker, New York.
[3] Alloway, J.A. and Raghavachari, M. (1991). Control charts based on Hodges-Lehmann estimator. Journal of Quality Technology,
23(4): 336–347.
[4] Alwan, L.C. (1995). The Problem of Misplaced Control Limits. Journal of the Royal Statistical Society. Series C (Applied Statistics),
44(3): 269–278.
[5] Brent, R.P. (1973). Algorithms for Minimization without Derivatives. Prentice-Hall, New Jersey.
[6] Duncan, A.J. (1956). The Economic Design of barX Charts used to Maintain Current Control of a Process. Journal of the American
Statistical Association, 51(274): 228–242.
[7] Ðorić, D., Mališić, J., Jevremović, V. and Nikolić-Ðorić, E. (2007). Atlas raspodela. Gradjevinski fakultet, Beograd.
[8] Gil-Pelaez, J. (1951). Note on the inversion theorem, Biometrika 38, pp. 481 - 482.
[9] Gupta, S.S. and Han, S. (1992). Selection and ranking procedures for logistic populations, In: Order Statistics and Nonparamet-
rics:Theory and Applications (Edited by P.K. Sen and I.A. Salama), 377–404. Elsevier, Amsterdam,
[10] Hill, I.D., Hill, R. and Holder, R.L. (1976). Algorithm AS 99: Fitting Johnson Curves by Moments. Journal of the Royal Statistical
Society - Series C (Applied Statistics), 25(2): 180–189.
[11] Janacek, G. J. and Meikle, S.E. (1997). Control charts based on medians. The Statistician, 46 (1): 19–31.
[12] Johnson, N.L. (1949). Systems of frequency curves generated by methods of translation. Biometrika, 36(1/2): 149–176.
[13] Johnson, N.L., Kotz, S. and Balakrishnan, N. (1994). Continuous Univariate Distributions Volume 1, second edition. Wiley, New
York.
[14] Johnson, N.L., Kotz, S. and Balakrishnan, N. (1995). Continuous Univariate Distributions Volume 2, second edition. Wiley, New
York.
[15] Li, D-C. and Lin, L-S. (2013). A new approach to assess product lifetime performance for small data sets. European Journal of
Operational Research, 230: 290–298.
[16] Kotz, S., Kozubowski, T.J. and Podgórski, K. (2001). The Laplace Distribution and Generalizations : a Revisit with Applications to
Communications, Economics, Engineering, and Finance. Springer Science+Business Media, New York.
[17] Küchler, U. and Tappe, S. (2008). On the shapes of bilateral Gamma densities. Statistics & Probability Letters 78(15): 2478ÂŰ-2484.
[18] Lorenzen, T.J. and Vance L.C. (1986). The Economic Design of Control Charts: A Unified Approach. Technometrics, 28(1): 3–10.
[19] Montgomery, D.C. (2005). Introduction to Statistical Quality Control, fifth edition. Wiley, New York.
[20] Saniga, E.M. (1989). Economic Statistical Control-Chart Design with an Application to X and R Charts. Technometrics, 31(3):
313–320.
[21] Scrucca, L. (2013). GA: A Package for Genetic Algorithms in R. Journal of Statistical Software, 53(4): 1–37.
[22] Witkovský, V. (2001). On the exact computation of the density and of the quantiles of linear combinations of t and F random
variables. Journal of Statistical Planning and Inference 94(1): 1–13.
 K.Veljkovic et al. / Filomat 29:10 (2015), 2325–2338 2338

[23] Witkovský, V. (2004). Matlab algorithm tdist: the distribution of a linear combination of StudentŠs t random variables.
COMPSTAT 2004 Symposium.
[24] Witkovsky, V. and Savin, A. (2005). tdist: Distribution of a linear combination of independent Student’s t-variables, R package
version 0.1.1.
[25] Woodall, W.H. (1985). The Statistical Design of Quality Control Charts. Journal of the Royal Statistical Society Series D 34(2):155–160.
[26] Wu, H. and Govindaraju, K. (2014). Computer-aided variables sampling inspection plans for compositional proportions and
measurement error adjustment. Computers and Industrial Engineering, 72: 239–246.
[27] Wu, J-W., Hong, C-W. and Lee, W-C. (2014). Computational procedure of lifetime performance index of products for the Burr XII
distribution with upper record values. Applied Mathematics and Computation, 227: 701–716.