Proceedings of the 2012 International Conference on Industrial Engineering and Operations Management

Istanbul, Turkey, July 3 6, 2012

580

T
2
Based Methods for Monitoring Gamma Profiles

Amirhossein Amiri and Mehdi Koosha
Department of Industrial Engineering
Shahed University, Tehran, Iran

Armaghan Azhdari
Department of Industrial Engineering
PNU, Tehran, Iran

Abstract

Profile is a relationship between a response variable and one or more explanatory variables that should be monitored
over time. There are many researches in this area but in most of them, it is assumed that the response variable is
normally distributed. However, sometimes this relationship is better characterized by Generalized Linear Models.
Yeh et al. (2009) proposed five T
2
based methods for monitoring of logistic regression profiles in which the response
variable follows Bernoulli distribution. This paper evaluates the performance of the T
2
methods for monitoring
Gamma response profiles in Phase I via simulation studies. The evaluation is done under different magnitudes of
step shifts and drifts in terms of the signal probability.

Keywords
Statistical process control; Profile monitoring; T
2
control chart; Gamma profile; Phase I

1. Introduction
In many real case problems, the relationship between response variable and explanatory variables which is called as
profile is desirable to be monitored over time instead of the response variable itself. According to the type of this
relationship, profiles are classified into categories such as Simple linear profiles, multiple linear profiles, Polynomial
profiles, multivariate linear profiles, non-linear profiles, Logistic profiles, and so on. Many studies have been done
by researchers for monitoring different types of profiles. For example Kang and Albin (2000), Kim et al. (2003),
Mahmoud and Woodal (2004), Mahmoud et al. (2007), Saghaei et al. (2009) Chen and Nembhard (2011) and
Noorossana et al. (2004) have proposed some methods for monitoring simple linear profiles.

Zou et al. (2007), and Mahmoud (2008) proposed some approaches for monitoring multiple linear profiles. Amiri et
al. (2010) and Kazemzadeh et al. (2008) have contributed in the area of polynomial regression profiles. Eyvazian et
al. (2011) and Noorossana et al. (2010) have proposed methods on monitoring multivariate linear profiles. Williams et
al. (2007) and Vaghefi et al. (2009) developed some methods for monitoring nonlinear profiles in Phases I and II,
respectively.

In all of the aforementioned researches, distribution of the response variable is assumed to be normal. However, in
some real case problems, response variable follows other Exponential family distributions such as Bernoulli, Poisson,
Exponential, Gamma, and etc. Yeh et al. (2009) proposed five T
2
based methods in order to monitor logistic profiles
in which the response variable follows Bernoulli distribution. Amiri et al. (2011) evaluated two of the best T
2
methods
proposed by Yeh et al. (2009) for monitoring Poisson regression profiles.

This paper concentrates on phase I monitoring of Gamma regression profiles. Gamma distribution is sum of m
identical independent exponential variables. There are many real cases in which the response variable follows Gamma
distribution. For example when each observation represents the time interval between the occurrence of two
successive defects in a process, the observations follow exponential distributions and if we consider the sum of these
exponential variables in each level of the explanatory variable as an observation, then each observation follows
Gamma distribution.

581

The rest of this paper is classified as follows: Section 2 proposes the procedure of parameter estimation in Gamma
regression profiles. The T
2
methods considered in this paper are explained in Section 3. Section 4 is dedicated to the
numerical example to evaluate the performance of the proposed methods. Our concluding remarks are given in
Section 5.
2. Gamma Regression Parameter Estimation
Suppose that there are m independent profiles (i=1, 2,., m) and the observations in each level of the explanatory
variable (j=1,2,,n) are repeated for k (k=1,2,,K) times. Vector of explanatory variables for each profile is as
follows:
x
i
= [x
i1
, x
i2
,., x
in
]
Let Z
ijk
be the k
th
observation of j
th
setting level in i
th
profile which is assumed to follow an exponential distribution
with mean
1
ij

. We define Y as sum of K

observations in j
th
level of i
th
profile. Hence,
,
1 =
=

K
ij ikj
k
Y Z (1)
As Y
ij
is sum of K exponential observations, it follows Gamma distribution. So the density probability function of
Y
ij
is as follows:
( )
1
( )

ij ij
ij
K
y ij
K
Y ij ij
f y y e
K

(2)
Hence, ( ) =
ij
ij
K
E Y

and
2
var( ) =
ij
ij
K
Y

. In addition, let
0 1
= +
ij ij
x .
We use Log link function (g) to make a relationship between
ij
and
ij
. Hence,
( )
( ) log ( ) = = = +
ij ij ij ij
g E Y x
0 1
, (3)
consequently,
( )
exp
=
ij
ij
K

,
(4)
where
0
and
1
are the regression parameters of each profile.

Since Y
ij
in each experimental setting level is independent to other levels observations, the joint likelihood function
of Y
i1
,Y
i2
,.,Y
in
is as follows:
1
1 1
1
( , )
[( 1)!]
=

=
=
=

n
ij ij
j
n
K
ij y n j K
ij ij ij
n j
L y y e
K

(5)
By taking logarithm of both sides of Eq. (5) and replacing
ij
with
ij
, the following equation is obtained:
( )
( )
( )
1
1 1
log[ ( , )] log( )
exp
( 1) log( ) log 1 !
exp
n
ij ij
ij
j
n n
ij
ij
ij
j j
K
L y K
y
K y K n K
=
= =
= +

(6)
By derivation of Eq. (6) with respect to , we obtain the score function in Eq. (7).
1 1
log[ ( , )]

= =

= +

n n
ij ij
ij ij ij ij
j j
L y
K x y x

(7)
For obtaining the estimations of parameters, we should put the computed score function equal to zero and multiply
both sides to
( )
exp
ij
K

.
So the Maximum Likelihood Estimator (MLE) of is the result of the following score function:
582

( ) ,
T
i i i
X = Y 0 (8)
where ( ) =
i i
exp ,
, , , , , , , ,
i
i
X
x
(
=
(

1 1 1 1 1 1 1 1 1
, 0=
T
) 0 ,..., 0 ( ,
i
= ( ,...., )
i in

1
and Y
i
= ( ,...., )
i in
Y Y
1
. By applying the
iterative weighted least square estimation method which is described below, the MLE estimators of the regression
parameters can be obtained. The estimation procedure of Gamma regression parameters are explained as follows:

Step 1: Set L=0. Consider
) 0 (

as an initial estimate of that can be estimated by any method such as MLE.

Step 2: Compute
) (L
i
by using Eq. (3). Then compute the vector of
i
as follows:
( )
( )
.
exp( )
L
i L
i
K
=

(9)
Note that Eq. (9) is derived from Log link function.
Step 3: Define a diagonal matrix of W by using Eq. (10).
( )
( )
( )
L
L
i
K
W diag
| |
|
=
|
|
\ .

2

(10)
So W is a nn diagonal matrix which its main diameters elements are
( )
( ) L
i
K

2
.
Step 4: Compute the adjusted dependent vector of
) (L
i
q by Eq. (11).
) ( ) (
1 ) ( ) ( ) (
i i
L L
i
L
i
y W q + =

(11)
Step 5: Update the estimation of by Eq. (12)
( ) ( ) ( ) ( )

( )
L T L T L L
i i
X W X X W q
+
=
1 1
(12)
if
( ) ( )

L L

+
<
1
, then
( )

+1
can be considered as the final estimation of . Otherwise set L=L+1 and go to step
2. is a predetermined small number (for example 0.005).

3. T
2
Methods
As mentioned former, Yeh et al. (2009) proposed five T
2
based methods in order to monitor logistic profiles. We
apply these methods for the case that the observations follow Gamma distribution. The methods will be described
below:
3.1 T
2
based on sample mean and covariance matrix (
2
H
T )
In this method, the covariance matrix is estimated as follows:
( )( )

,
m
H i i
i
S
m

=

1
1
1

(13)
where

i
is vector of estimated parameters for i
th
profile and is the average of estimated vectors among all m
profiles. The T
2
statistic of this method can be computed as Eq. (14).
,

( ) ( )
H i i H i
T S

=
2 1
, (14)
3.2 T
2
based on the sample average and intra-profile pooling (
I
T
2
)
The statistic of this method is shown in Eq. (15).
2 1
,

( ) ( )
T
I i i I i
T S

= , (15)
where

=

=
= =
m
i
i
m
i
i I
X W X
m m
S
1
1
1
) (
1
)

var(
1

583

3.3 T
2
based on sample average and moving ranges (
R
T
2
)
In this method, the T
2
statistic that should be mapped on the control chart is shown in Eq. (16)
2 1 T
R,i i R i

T ( ) S ( )

= , (16)
where
1
1 1
1
1

( )( )
2( 1)
m
T
R i i i i
i
S
m

+ +
=
=

3.4 T
2
based on Minimum Volume Ellipsoid (
MVE
T
2
)
For this method, the T
2
statistic is given in Eq. (17)
2 1
,

( ) ( )
E i i E E i E
T S

= ,

(17)
where
E
and S
E
are estimated by the MVE method.

3.5 T
2
based on minimum covariance determinant (
MCD
T
2
)
By applying Eq. (18), one can compute theT
2
statistic for this method.
2 1
,

( ) ( )
D i i D D i D
T S

=
(18)

D
and S
D
are estimated by the MCD method.
Details about the algorithms of MVE and MCD are presented in Vargas (2003).
4. Performance Evaluation of the Methods
In this section, the performance of the methods in detecting step shifts and drifts is evaluated using simulation studies.
The vector of explanatory variable is adopted from Yeh et al. (2009). So:
X [Log(0.1), Log(0.2),....., Log(0.9)] =

0
and
1
are assumed to be 3 and 2, respectively when the profiles are under statistical control. m and K are
considered to be equal to 30. We consider two kinds of shifts including step shifts and drifts in this paper. So, we first
explain the concepts of step shifts and drifts.
4.1 Step shifts
In this type of shifts, the mean of the regression parameters is in statistical control up to a special profile. After that
profile, the mean of the regression parameters changes to another value and remains in that level up to a time a
corrective action is implemented. Hence,
1 , 0
1 3 + =
out
and
2 , 1
3 2 + =
out

1
and
2
can be computed as follows (Yeh 2009):
|
|
.
|

\
|
= =
|
|
.
|

\
|

1595 . 0 0617 . 0
0617 . 0 0413 . 0
) (
1
2
2 2 1
2 1
2
1
WX X
T



In this paper we have applied the step shifts on the second half of profiles (16
th
to 30
th
profiles).
4.2 Drifts
In this type of shift, the regression parameters are in statistical control and from a special point, the mean of the
regression parameters changes with a linear trend. So the magnitude of shift increases by incensing in the sample
number. For example (1,3) means that:
1 , 0
1
1
1
3 |
.
|

\
|

+ =
m
i
out
and
2 , 1
3
1
1
2 |
.
|

\
|

+ =
m
i
out

Drifts are implemented from the second profile in this paper.

Performance of the methods under various step shifts and drifts are shown in Tables 1 and 2, respectively. Note that
we use power of the control chart criterion which determines the probability of detecting each shift in 10000
replications as the basis of our comparisons.
584




Table 1: Performance of the T
2
methods under step shift

Methods

2
I
T
2
R
T
2
H
T
2
MVE
T
2
MCD
T
UCL
Shifts
40.5000

12.3944

10.5336

18.0500

58.0004
0,0

0.0574 0.0508 0.0536 0.0507 0.0523
1,1

0.0624 0.5358 0.0288 0.0283 0.0533
1,0

0.1282 0.1200 0.0490 0.0387 0.0540
1,2

0.0550 0.9694 0.0240 0.0370 0.0517
2,3

0.0560 0.9972 0.0220 0.0323 0.0563
2,0

0.3832 0.3316 0.0300 0.0353 0.0399
1,3

0.0772 0.9984 0.0228 0.0278 0.0382
2.5,0

0.5760 0.4784 0.0270 0.0397 0.0419
3,0.3

0.6862 0.4566 0.0290 0.0340 0.0536
3,0

0.7588 0.6107 0.0256 0.0343 0.0545

Table 2: Performance of the T
2
methods under drift

Methods

2
I
T
2
R
T
2
H
T
2
MVE
T
2
MCD
T
UCL
Shifts
40.5000 12.3944 10.5336 18.0500
58.0004
0,0

0.0574 0.0508 0.0536 0.0507 0.0523
1,1

0.0570 0.3290 0.0422 0.0510 0.0529
1,0

0.0870 0.0822 0.0570 0.0547 0.0532
1,2

0.0526 0.9672 0.0264 0.0480 0.0519
2,3

0.0610 0.9996 0.0282 0.0530 0.0535
2,0

0.2128 0.1822 0.0492 0.0457 0.0499
1,3

0.0526 1.0000 0.0294 0.0557 0.0523
2.5,0

0.3054 0.2832 0.0420 0.0537 0.0517
3,0.3

0.3690 0.2378 0.0432 0.0383 0.0492
3,0

0.4188 0.3982 0.0428 0.0507 0.0511


Results show that
2
I
T performs better than the other T
2
based methods under individual step shift and drift in
0
.
Under simultaneous shifts in
0
and
1
, the
2
R
T method outperforms the other methods.

5. Conclusions
There are many real case problems in which the response variable of the profiles follows Gamma distribution. For
example when each observation describes the interval between the occurrence of two defects, the sum of k
observations for each level of the explanatory variable follows Gamma distribution. In this paper, we extend and
evaluate the performance of the five T
2
methods proposed by Yeh et al. (2009) for monitoring Gamma profiles. As a
matter of fact, the procedure of parameter estimation for Gamma profiles is the other contribution of this paper.
585

Results showed that the
2
I
T method performs better than the other T
2
methods in the case of individual step shifts and
drifts in
0
. However,
2
R
T performs better than the other T
2
methods when simultaneous shifts, either step shifts or
drifts, occur in the regression parameters
0
and
1
.

References
Amiri A., Jensen W. A., Kazemzadeh R. B., A Case Study on Monitoring Polynomial Profiles in the Automotive
Industry, Quality and Reliability Engineering International, vol. 26, pp. 509-520, 2010.
Amiri, A., Koosha, M., Azhdari, A., Profile monitoring for Poisson response, Proceedings of IEEM conference,
IEEE, Singapore, 2011.
Chen, S., Nembhard H. B., A High Dimensional Control Chart for Profile Monitoring, Quality and Reliability
Engineering International, vol. 27, pp. 451-464, 2011.
Eyvazian M, Noorossana R, Saghaei A., Amiri A., Phase II Monitoring of Multivariate Multiple Linear Regression
Profiles, Quality and Reliability Engineering International, vol. 27, pp. 281-296, 2011.
Kang, L., Albin, S. L., On-Line Monitoring when the Process yields a Linear Profile, Journal of Quality
Technology, vol. 32, pp. 418-426, 2000.
Kazemzadeh, R. B., Noorossana, R., Amiri, A., Phase I Monitoring of Polynomial Profiles, Communications in
Statistics: Theory and Methods, vol. 37, pp. 1671-1686, 2008.
Kim, K., Mahmoud, M. A., Woodall, W. H., On the Monitoring of Linear Profiles, Journal of Quality Technology,
vol. 35, pp. 317-328, 2003.
Mahmoud, M. A., Phase I Analysis of Multiple Linear Regression Profiles, Communications in Statistics:
Simulation and Computation, vol. 37, pp. 2106-2130, 2008.
Mahmoud, M. A., Parker, P. A., Woodall, W. H., Hawkins, D. M., A Change Point Method for Linear Profile Data
Quality and Reliability Engineering International, vol. 23, pp. 247-268, 2007.
Mahmoud, M. A., Woodall, W. H., Phase I Analysis of Linear Profiles with Calibration Applications,
Technometrics, vol. 46, pp. 380-391, 2004.
Noorossana, R., Amiri, A., Vaghefi, S.A. and Roghanian, E., .Monitoring Quality Characteristics Using Linear
Profile, Proceedings of the 3rd International IndustrialEngineering Conference, Tehran, Iran, pp. 246-255,
2004.
Noorossana, R., Eyvazian M., Amiri A., Mahmoud M.A., Statistical Monitoring of Multivariate Multiple Linear
Regression Profiles in Phase I With Calibration Application, Quality and Reliability Engineering International
vol. 26, pp. 291-303, 2010.
Saghaei, A., Mehrjoo, M., Amiri, A., A CUSUM-based Method for Monitoring Simple Linear Profiles,
International Journal of Advanced Manufacturing Technology, vol. 45, pp. 1252-1260, 2009.
Vaghefi, A., Tajbakhsh, S. D., Noorossana, R., Phase II Monitoring of Nonlinear Profiles, Communications in
Statistics: Theory and Methods, vol. 38, pp. 1834-1851, 2009.
Vargas J. A., Robust Estimation in Multivariate Control Charts for Individual Observation, Journal of Quality
Technology, vol. 35, pp. 367-376, 2003.
Williams, J. D., Woodall, W. H., Birch, J. B., Statistical Monitoring of Nonlinear Product and Process Quality
Profiles, Quality and Reliability Engineering International, vol. 23, pp. 925-941, 2007.
Yeh, A. B., Huwang, L., Li, Y. M., Profile Monitoring for a Binary Response, IIE Transactions, vol. 41, pp. 931-
941, 2009.
Zou, C., Tsung, F., Wang, Z., Monitoring General Linear Profiles using Multivariate Exponentially Weighted
Moving Average schemes, Technometrics, vol. 49, pp. 395-408, 2007.