Professional Documents
Culture Documents
www.elsevier.com/locate/csda
The quality control chart for monitoring multivariate
autocorrelated processes
Jeffrey E. Jarrett
a,
, Xia Pan
b
a
University of Rhode Island, USA
b
University of Illinois at Springeld, USA
Received 13 July 2005; received in revised form 15 January 2006; accepted 17 January 2006
Available online 30 March 2006
Abstract
Previously, quality control and improvement researchers discussed multivariate control charts for independent processes and
univariate control charts for autocorrelated processes separately. We combine the two topics and propose vector autoregressive
(VAR) control charts for multivariate autocorrelated processes. In addition, we estimate AR(p) models instead of ARMA models for
the systematic cause of variation. We discuss the procedures to construct the VAR chart. We examine the effects of parameter shifts
and by example present procedures to show the feasibility of VAR control charts. We simulate the average run length to assess the
performance of the chart.
2006 Elsevier B.V. All rights reserved.
Keywords: Quality control and improvement; VAR control charts; Multivariate autocorrelated processes
1. Introduction and purpose
Shewhart (1931) developed statistical process control (SPC) techniques for the improvement and assurance of product
and service quality. The assumptions of these methods include serial independence of successive samples taken from
a process. Alwan and Roberts (1988) proposed modeling and ltering to monitor residuals by a one-step-forecast of
autocorrelated process possessing systematic patterns. English and Sastri (1990) applied Kalman lter to monitoring
univariate autoregressive (AR) (1) process. Wardell et al. (1994) derived the run length distribution of Shewhart
chart on residuals. Montgomery and Mastrangelo (1991), Lu and Reynolds (1999) discussed some SPC charts for
autocorrelated data. Runger and Willemain (1995) compared Shewhart charts based on several process transformations
to eliminate autocorrelation. Snoussi et al. (2005) indicated that residuals control charts provide much better shift
detection properties. Testik (2005) considered using an AR (1) process to modify control limits for residual based
charts.
Loredo et al. (2002) provided a method for monitoring autocorrelated processes based on Hawkins (1991, 1993)
regression adjustment. They examined the performance of residual based quality control charts in terms of the average
run length to observation based control charts via Monte Carlo simulations. Their results indicate the superiority of
residual based control in comparison with the observation based control charts when detecting a mean shift in short run
1
1
L + 1
2
L
2
+ + 1
p
L
p
y
t
+ e
t
= c + 1(L) y
t
+ e
t
, (1)
where L is the backshift operator, c = (c
1
, c
2
, . . . , c
n
)
is the vector
for the error term. Each 1
j
is a (n n) coefcient matrix for lag j. e
t
contain no autocorrelation but does contain
cross-sectional correlation. That is, E
e
t
e
1
1
1
2
1
p
t
=
1, y
t 1
, y
t 2
, . . . , y
t p
being a 1(1+np)
vector. Then Eq. (1) becomes
y
t
= H
x
t
+ e
t
. (1
)
Due to the usual assumptions about e
t
, we estimate the coefcient matrix n
0, O
t =1
y
t
x
t =1
x
t
x
1
, (2)
where T is the number of observations in Stage I. The estimated systematic model is
y
t
=
H
x
t
. (3)
In addition, we estimate error term as
e
t
= y
t
y
t
. We then develop the special-cause control chart based on e
t
.
Maximum likelihood methods estimate the covariance of e
t
as
O = (1/T )
T
t =1
e
t
e
t
p
O. (4)
If the process is in-control, the model in (1) is appropriate and the error termis uncorrelated and normal, the estimated
error e
t
should be also approximately normal and invariant with time. Thus, it is asymptotically N
0, O
. We now
apply the multivariate control chart e
t
. In Stage I the Hotelling T
2
statistic is, approximately ,
2
(n) when T n is large.
T
2
=
e
t
O
1
e
t
,
2
(n), (5)
where
e is mean of Stage I,
e =
T
t =1
e
t
/T , which should be close to zero. Since the estimates of the AR(p) model are
asymptotically available, we use chi-square instead of Beta distribution.
For Stage II observations, Y
i
, (i >T ), we denote the rst Stage II observation as the (T + 1)st observation. We
substitute the estimated error by the one-step-ahead forecast error, while we compare the rst Stage II observation with
the one-step-ahead forecast from Stage I. That is, from the known T observations in Stage I, we predict the rst Stage
II status as
y
T +1
=
H
x
T +1
. (6)
Moreover, the forecasting error is the difference between the real observation and the forecast,
e
T +1
=
y
T +1
y
T +1
. (7)
There are several possible schemes for Stage II chart. If we treat
O as Wishart distribution, the chart statistic is
T
2
T +1
=
e
T +1
O
1
e
T +1
n(T + 1)(T 1)
T (T n)
F
:,n,T n
. (8)
However, in an autoregression environment, we only have asymptotic properties,
O O.
Therefore, the more appropriate way is to use the approximation:
T
2
T +1
=
e
T +1
O
1
e
T +1
,
2
(n). (9)
J.E. Jarrett, X. Pan / Computational Statistics & Data Analysis 51 (2007) 38623870 3865
For the following observations, the forecasting errors are acquired recursively one-step-ahead each. Generally, for
the forecasting error of the ith observation,
e
i
=
H
x
i
y
i
, the chart statistic is
T
2
i
=
e
i
O
1
e
i
,
2
(n). (10)
An F distribution is asymptotically close to a chi-square distribution (divided by T ). The asymptotic distributions of
the estimates provide, under the same condence level, more narrow condent intervals for true parameters than the
exact distribution for limited sample size T. Therefore, control charts using chi-square distributions are more sensitive
than those using F distributions.
Selecting statistic: One statistic is given in (10). By selecting the expression of the mean itemof T
2
i
we can have more
chart schemes. One option is directly choosing
O
1
e
i
,
2
(n). (11)
The corresponding Stage I statistic is
T
2
=
e
t
O
1
e
t
,
2
(n). (5
)
Another option is using recursive mean. This is corresponding to the chart of moving centerline in univariate case
(Montgomery and Mastrangelo, 1991). The mean for the (i + 1)th observations statistic can be dened as
e
i
=
e
i
+ (i 1)
e
i1
e
i
e
i
O
1
e
i
e
i
,
2
(n). (12)
Although the recursive mean varies every time when a new observation is added, the control limit of the chart does
not change. This is much better than the moving center line in the univariate case. Note, all the above schemes have
the same chart that we determine the control limit (UCL) by ,
2
(n) say, ,
2
0.01
(n). Only the statistic values are different
for different schemes. Hence, the VAR chart is actually simple in its demonstration, and model estimation is relatively
simple. Among (10)(12), the statistic in (11) is the most convenient. Because the asymptotic mean of
e
t
is zero thus
(11) is theoretically justied for large sample size.
Determining the AR order: In practice, determining the autocorrelation order p may be the issue before chart setting.
To nd a parsimonious AR order, the null hypothesis is that the AR order is p
0
as opposed to being p
1
>p
0
. The test
statistic is a likelihood ratio:
2 (L
1
L
0
) = T
log
O
0
log
O
1
,
2
n
2
(p
1
p
0
)
, (13)
where T is the sample size in Stage I. O
0
is the estimated covariance matrix of the error term given in (10) under the
null hypothesis being true. O
1
is also from (10) but under the alternative hypothesis. The chi-square distribution has a
degree of freedom n
2
(p
1
p
0
) because each variable has (p
1
p
0
) less lags under the null hypothesis and each of
the n equations include n variables and totally n
2
restrictions.
3. The effect of shift
There are three classes of parameters in VAR models: the process mean j, the shift on covariance matrix of error
term, O, and the shift autocorrelation structure of the model, 1. The most important shift of interest is the shift on the
mean j of the process.
For a mean shift, the relationship between the mean of a stationary time series vector, j is the expectation of Eq. (1).
The following theorem gives the effects of mean shift on the VAR chart statistic. There are possibly two ways for a shift
to build on the ARMA process. One is that the shift is added on the constant item, meaning there is really a physical
3866 J.E. Jarrett, X. Pan / Computational Statistics & Data Analysis 51 (2007) 38623870
structure that the historical statuses are added on the constant. Another way is that the shift is added on the mean of the
process, which is more an observed result. We here only consider the rst situation.
Theorem 1. For a process model in (1) and VAR chart in (10), a shift
o in the process constant adds a normal
distribution with mean of
n
O
1
n
k
and variance of 4
k
O
1
n
k
on the in-control T
2
. The effects depend on the shift
magnitude and the estimates of VAR parameter matrices, n = (I
n
1(L))
o.
Proof. In the following proof, we use the estimated coefcients rather than the known coefcients. Denote the in-
control residual here as
e
t
, and the out-of-control (due to shift) residuals as
c
t
. In addition, we use z
t
to denote the
process after shift. These changes in denotation are not necessary but just for an alternative expression. The results
for the shift effects are the same except that one is univariate ARMA and the other is multivariate AR, because the
asymptotic estimates should be just the known parameters.
The forecasting error without shift occurrence is
e
t
= y
t
y
t
. After the shift, the process status is z
t
rather than y
t
.
A sudden shift at time just before t
0
on mean for some of the variables implies
z
t
= y
t
+
o for t t
0
.
Since the shift is unknown before detection, we still base the one-step-ahead forecasting on the estimated coefcient
matrix
H
c
1
z
t
=
c +
1(L) z
t
when t t
0
+ p.
Then the forecasting error after the shift is
t
t
= z
t
z
t
= z
t
c
1(L) z
t
= y
t
+
o
c
1(L) y
t
1(L)
o
= y
t
1(L)
o
y
t
+
o =
e +
I
n
1(L)
o
=
e
t
+
n.
Here the
e
t
was the forecasting error at time t, had the shift not occurred. With the shifted
t
=
e
t
n, the Hotelling
T
2
statistic will be a noncentral chi-square distribution:
shift
T
2
t
=
t
O
1
t
=
e
t
n
O
1
e
t
n
=
e
t
O
1
e
t
2
O
1
e
t
+
n
O
1
n
=
noshift
T
2
t
2
O
1
e
t
+
n
O
1
n.
The rst term is the in-control Hotelling statistic
noshift
T
2
t
,
2
(n) with mean of n and variance of 2n. Since
asymptotically
e
t
N
0, O
O
1
e
t
is a linear combination of normal distributions so that it
is normal with mean of 0 and variance of 4
O
1
O
O
1
n = 4
O
1
O
1
n and variance of 4
O
1
n on T
2
by
n
O
1
n.
When t
0
t <t
0
+ p, let k = t t
0
. We have z
t
= y
t
+
o for t
0
t <k and k lags shift while p k lags do not shift.
The one-step-ahead prediction is
z
t
=
c +
1
1
z
t 1
+
1
2
z
t 2
+ +
1
k
z
t k
+
1
k+1
y
t k1
+ +
1
p
y
t p
=
c +
1
1
L +
1
2
L
2
+ +
1
k
L
k
z
t
+
1
k+1
y
t k1
+ +
1
p
y
t p
.
J.E. Jarrett, X. Pan / Computational Statistics & Data Analysis 51 (2007) 38623870 3867
Thus
t
t
= z
t
z
t
= z
t
c +
1
1
L +
1
2
L
2
+ +
1
k
L
k
z
t
+
1
k+1
y
t k1
+ +
1
p
y
t p
= y
t
+
o
c +
1
1
L +
1
2
L
2
+ +
1
k
L
k
y
t
+
1
k+1
y
t k1
+ +
1
p
y
t p
1
1
L +
1
2
L
2
+ +
1
k
L
k
o
=
e
t
+
n
k
,
where
n
k
=
I
n
1
1
1
2
1
k
o and k = 0, 1, 2, . . . , (t t
0
).
Therefore the effect on VAR chart statistic is, for t
0
t t
0
+ p,
shift
T
2
t
=
noshift
T
2
t
2
O
1
e
t
+
n
O
1
n
k
,
where k =t t
0
. This is still a positive shift on the mean of the in-control T
2
and on the variance of the in-control T
2
.
From the denition of
n and
n
k
, obviously
n =
n
p
for t t
0
+ p, because 1
k
= 0 for k >p, and the estimates of 1
k
should have the same property asymptotically. The effects during t
0
t t
0
+p can be either larger or smaller than the
effects during t t
0
+ p, depending on the specic VAR structure
1(L) and
O.
Other systematic shifts, such as the shift on the covariance matrix of error term, O, or the shift on the autocorrelation
structure of the model, 1, will also violate the distributions of the chart statistic, hence out of control signals will be
observed after the shifts. The larger the shift scale, the more serious the violation of the chi-square distribution. The
out-of-control ARL, which determines when the VAR chart detects the shift, relates to the control limit and the scale
of the shift.
4. A simulated example of a VAR chart
Following is an illustration showing the procedures and properties of a VAR chart. We specify three as both the order
of autocorrelation and the number of variables. We design the covariance matrix of the error term vector as
O =
1 0.5 1.5
0.5 4.25 1.75
1.5 1.75 2.99
.
In this way, we keep the positive denite. The constant term is set to be (3
, 1
2
=
, 1
3
=
.
By this design, all of the characteristic values of
I
n
z
p
1
1
z
p1
1
2
z
p2
1
p
= c
I
n
1
1
L 1
2
L
2
1
3
L
3