STATGRAPHICS – Rev.
7/24/2009
Multivariate EWMA Control Chart
Summary
The Multivariate EWMA Control Chart procedure creates control charts for two or more
numeric variables. Examining the variables in a multivariate sense is extremely important when
the variables are highly correlated, since joint out-of-control conditions can occur without any
individual variable violating its control limits when plotted separately. The EWMA chart is
similar to the T-Squared chart, except that the points plotted on the chart are a weighted average
of current and past observations.
Sample StatFolio: mvewma.sgp
Sample Data:
The file grit.sgd contains measurements made on n = 56 batches of “grit”, from Holmes and
Mergen (1993). The data represent the percentages of large, medium, and small particles in the
grit. The table below shows a partial list of the data in that file:
Large Medium Small Means Covariances
5.4 93.6 1.0 5 3.5
3.2 92.6 4.2 90 -5.5
5.2 91.7 3.1 -5.5
3.5 86.9 9.6 13.5
2.9 90.4 6.7
4.6 92.1 3.3
4.4 91.5 4.1
5,0 90.3 4.7
8.4 85.1 6.5
4.2 89.7 6.1
3.8 92.5 3.7
4.3 91.8 3.9
3.7 91.7 4.6
3.8 90.3 5.9
2.6 94.5 2.9
Since the percentages in the first three columns sum to 100% in each row, it is only necessary to
create a chart based on the first 2 columns. In addition to the data, the file contains columns with
the standard means and covariances for the percentage of large and medium particles, established
when monitoring of the process was first begun.
2009 by StatPoint Technologies, Inc. Multivariate EWMA Control Chart - 1
STATGRAPHICS – Rev. 7/24/2009
Data Input
The data to be analyzed consist of p = 2 or more numeric columns containing the variables of
interest, with optional entry of the variable means and variance-covariance matrix.
Data: 2 or more numeric columns containing the n samples, one sample per row.
Subgroup numbers or size: If the data were obtained as individuals, leave this field blank or
enter “1”. If the data were collected in subgroups, each of size m, enter the single value m. In
such a case, each consecutive m rows in the file will be considered to form a subgroup. If the
subgroup sizes are not equal, enter the name of an additional numeric or non-numeric column
containing group identifiers. The program will scan this column and place all sequential rows
with identical codes into the same group.
Standard Means: For an “initial study” or Phase I analysis where the data will be used to
determine the control limits, leave this field blank. For a “control-to-standard” or Phase II
analysis, enter the name of a column contain p means.
Standard Covariances: For an “initial study” or Phase I analysis, leave this field blank. For
a “control-to-standard” or Phase II analysis, enter the name of a column contain the p2
variances and covariances. In entering the values in a covariance matrix, enter the values in
the first row of the matrix, then the values in the second row, and so forth. Note: if you select
Save Results after performing a Phase I analysis, the covariance matrix will be saved in this
exact format.
2009 by StatPoint Technologies, Inc. Multivariate EWMA Control Chart - 2
STATGRAPHICS – Rev. 7/24/2009
Labels: optional labels for each subgroup. The labels will be applied in sequence to the
subgroups when plotting the control charts.
Select: subset selection.
Analysis Summary
The Analysis Summary shows the number of observations included in the analysis and the
location of the control limits on the control charts.
Multivariate EWMA Charts
Data variables:
Large
Medium
Number of observations included = 56
Number of observations excluded = 0
Smoothing parameter lambda: 0.2
Initialization: centerline
Phase 2 - covariance specified based on standard
Chart Alpha LCL UCL Beyond limits
T-Squared 0.0027 0.0 11.829 15
Included in the table are:
Smoothing parameter lambda: the value of the EWMA parameter , specified on the
Analysis Options dialog box. The default value of is determined from the settings on
the Control Charts tab of the Preferences dialog box, accessible from the Edit menu.
Phase: If Phase 1, the method for estimating the covariance matrix is displayed. If Phase
2, the assumptions about the input covariance matrix are shown.
Chart: the type of chart. For individuals data, only a T-Squared chart is created. For
grouped data, a generalized variance chart is included if the subgroup size exceeds the
number of variables.
Alpha: the false alarm probability of the chart, specified using Analysis Options. For
standard 3-sigma control charts, = 0.0027.
LCL: the lower control limit.
UCL: the upper control limit.
Beyond limits: the number of points on the control chart that are beyond the control
limits.
In the example, the process generated 15 out-of-control signals on the EWMA T-Squared chart.
2009 by StatPoint Technologies, Inc. Multivariate EWMA Control Chart - 3
STATGRAPHICS – Rev. 7/24/2009
Analysis Options
Alpha: the false alarm probability for positioning the control limits. For the equivalent of a
standard 3-sigma control chart, set = 0.27%.
Covariance Matrix: details about the covariance matrix. There are 4 possibilities:
1. If the Standard Covariances field was left blank on the data input dialog box and the data
are in subgroups, no entry is necessary since the covariances will be estimated from the
pooled within-group variability.
2. If the Standard Covariances field was left blank on the data input dialog box and the data
are individuals, select Pooled estimator to estimate the covariance between variables j
and k using the usual estimator
xij x j xik xk
1 n
s jk (1)
n 1 i 1
Select Successive differences to estimate the covariance using
xij xi1, j xik xi 1,k
n
1
s jk (2)
2(n 1) i 1
The second estimator is a more “local” estimator, in the sense that it captures only short-
term variability, in a similar manner to the way in which a moving range is used to
estimate variability for a standard individuals chart.
3. If an entry was made in the Standard Covariances field and the estimates provided were
obtained from a previous sample, enter the size of that previous sample (if grouped, the
number of subgroups) in the Standard Sample Size field.
4. If an entry was made in the Standard Covariances field and the covariances are assumed
to be known, leave the Standard Sample Size field empty.
Lambda: a value for the EWMA parameter 0 < < 1. The value of controls the amount of
weight given to the past history of the process. The smaller the value, the more weight given
to older observations or subgroups. This also impacts the average run length of the chart.
2009 by StatPoint Technologies, Inc. Multivariate EWMA Control Chart - 4
STATGRAPHICS – Rev. 7/24/2009
EWMA Chart
The EWMA Chart shows the exponentially weighted value of T2 for each data value or subgroup:
Multivariate EWMA Control Chart
UCL = 11.83, lambda = 0.2
24
20
T-Squared
16
12
8
4
0
0 10 20 30 40 50 60
Observation
The EWMA procedure begins by smoothing the observed data vector at time i by
EWMAi xi (1 ) EWMAi 1 (3)
for individuals data and by
EWMAi xi (1 ) EWMAi 1 (4)
for grouped data, with EWMA0 set equal to the mean vector or x . The i-th value of T-squared
is then calculated from
Ti 2 EWMAi Z1i EWMAi (5)
where is the covariance matrix of the input data, and
Zi
2
1 1
2i
(6)
For the sample data, the chart starts out well below the control limit but then rises during the
latter section. 15 out-of-control signals are generated by the chart.
2009 by StatPoint Technologies, Inc. Multivariate EWMA Control Chart - 5
STATGRAPHICS – Rev. 7/24/2009
Pane Options
Point Symbols: select Beyond Limits to draw special point symbols only for points falling
above the control limit. Select Largest Contributor to identify the variable that contributes
most to each value of T2.
Decimal Places for Limits: number of decimal places for displaying the control limit.
Color Zones: check this box to display green and red zones.
Example: Identifying Largest Contributor
If Largest Contributor is selected, the chart will take the following form:
Multivariate EWMA Control Chart
UCL = 11.83, lambda = 0.2
24 Largest
Large
20 Medium
T-Squared
16
12
8
4
0
0 10 20 30 40 50 60
Observation
Each point on the chart is coded according to the variable that contributes the most to the value
of T2. In the plot above, the biggest contributor to the first rise appears to be the percentage of
Medium particles, while the biggest contributor to the second rise appears to be the percentage of
Large particles.
2009 by StatPoint Technologies, Inc. Multivariate EWMA Control Chart - 6
STATGRAPHICS – Rev. 7/24/2009
Multivariate Control Chart Report
This pane tabulates the points on the control chart:
Multivariate Chart Report
Observation T-Squared Large Medium
27 * 12.3143 5.35513 86.8982
28 * 17.5104 5.4841 86.2185
29 * 17.9753 5.86728 85.6948
30 * 17.676 6.05382 85.5159
31 * 14.1056 6.10306 85.8327
44 * 11.9746 6.8905 85.8034
45 * 19.5133 7.6924 85.0827
46 * 20.3941 7.79392 86.0262
47 * 18.7157 7.57514 86.9009
48 * 14.9892 7.24011 87.5407
49 * 17.4784 7.53209 86.7526
50 * 14.3944 7.30567 87.0021
51 * 16.8681 7.52454 86.5417
52 * 22.2222 7.93963 85.3533
53 * 15.1529 7.3717 86.8827
An asterisk indicates any value beyond the control limits. Any points excluded from the analysis
using the Exclude button are indicated with an X.
Pane Options
Display All Subgroups – if checked, all observations or subgroups will be tabulated.
Otherwise, only points beyond the control limits will be included in the table.
Display Original Data – if checked, the values of each variable will be displayed.
Otherwise, only the control chart values will be tabulated.
2009 by StatPoint Technologies, Inc. Multivariate EWMA Control Chart - 7
STATGRAPHICS – Rev. 7/24/2009
Generalized Variance Chart
A T-Squared chart is designed to monitor the means of p variables. To monitor the variance, a
Generalized Variance Chart by sometimes be displayed:
Generalized Variance Chart
50
UCL = 43.0171
40 CL = 7.70563
Gen. Variance
LCL = 0.0
30
20
10
0
0 3 6 9 12 15
Subgroup
This chart is created only for data arranged in subgroups, and only if the subgroup size is at least
p + 1. The generalized variance |Si| for the i-th subgroup is defined as the determinant of its
variance-covariance matrix.
The above chart shows the grit data grouped in subgroups of 4 consecutive observations each.
Any point beyond the upper control limit would indicate an unusually large amount of variability
within that subgroup. In this case, no such points are present.
Control Ellipse
If out-of-control signals are shown on the control chart, it is useful to examine those values in
detail. A good chart to use in the case of p = 2 variables is the Control Ellipse:
Control Ellipse
95
93
Medium
91
89
87
85
2.8 3.8 4.8 5.8 6.8 7.8 8.8
Large
The upper control limit on the T-Squared chart corresponds to an elliptical region in the space of
any two of the variables, with the other variables held at a fixed value. For p = 2, any out-of-
control signals will show up as points outside the ellipse.
2009 by StatPoint Technologies, Inc. Multivariate EWMA Control Chart - 8
STATGRAPHICS – Rev. 7/24/2009
In the sample data, it will be noticed that some out-of-control singles correspond to a high
percentage of Large particles while others correspond to a low percentage of Medium particles.
Pane Options
Select 2 variables: select any 2 variables to define the control ellipse. The ellipse will be
plotted assuming that all other variables at held at their mean levels. Care should be taken in
interpreting the plot if p > 2, since the true elliptical control region for each point depends on
the value of the variables that are not plotted.
2009 by StatPoint Technologies, Inc. Multivariate EWMA Control Chart - 9
STATGRAPHICS – Rev. 7/24/2009
EWMA Decomposition
The T2 statistic can be decomposed into components attributable to each of the variables. One
method for measuring the contribution of the j-th variable to an out-of-control T2 value is by
looking at how much smaller T2 would be if the j-th variable was not included in the analysis.
The T-Squared Decomposition pane does such a decomposition for each out-of-control signal on
the T-Squared chart:
T-Squared Decomposition
Observation Large Medium
27 12.3143 5.90004 11.99
28 17.5104 7.97741 16.9078
29 17.9753 5.61899 16.0411
30 17.676 4.27107 14.8204
31 14.1056 2.52796 10.9768
44 11.9746 0.233548 2.78432
45 19.5133 3.39345 0.872939
46 20.3941 9.86659 0.321611
47 18.7157 12.3129 1.66376
48 14.9892 10.9573 2.08558
49 17.4784 10.448 0.991745
50 14.3944 8.40275 0.72444
51 16.8681 8.8947 0.479667
52 22.2222 7.8278 0.00137625
53 15.1529 8.67433 0.688634
Following Runger, Alt and Montgomery (1996), the table shows
d j T 2 T(2j ) (7)
where T(2j ) is the value of the statistic using all variables except the j-th.
For the current data, Medium appears to be the dominant variable for the early out-of-control
signals, while Large appears to be the dominant variable for the later signals.
3-D Control Chart
When the data consist of p = 3 variables, a 3-D control chart can be very helpful, since the
control region is then an ellipsoid.
2009 by StatPoint Technologies, Inc. Multivariate EWMA Control Chart - 10
STATGRAPHICS – Rev. 7/24/2009
Control Ellipsoid
2.4
1.4
X3
0.4
-0.6
1.2
0.8
-1.6 0 0.4
-1.2 -0.7 -0.2 -0.4
-0.8
0.3 0.8 -1.2
1.3 X2
X1
The above plot shows the outline of a typical control ellipsoid for 3 variables. Note: to explore
this plot, it is very helpful to use the dynamic rotation button on the analysis toolbar.
Pane Options
Select 3 variables: select any 3 variables to define the control ellipsoid. The ellipsoid will be
plotted assuming that all other variables at held at their mean levels. Care should be taken in
interpreting the plot if p > 3, since the true control region for each point depends on the value
of the variables that are not plotted.
2009 by StatPoint Technologies, Inc. Multivariate EWMA Control Chart - 11
STATGRAPHICS – Rev. 7/24/2009
Save Results
The following results can be saved to the datasheet:
1. T-Squared – the calculated T2 values for each observation or subgroup.
2. Means – the p variable means.
3. Covariances – the p2 variances and covariances in rowwise order.
4. Labels – the labels corresponding to each value of T2.
5. Generalized Variance - if calculated, the values of |S|.
Calculations
T-Squared Control Limit if Covariances are Known
UCL X 2, p (8)
T-Squared Control Limit if Covariances are Estimated from k Previous Samples
p (k 1)(k 1)
UCL F , p ,k p (9)
k (k p)
T-Squared Control Limit if Covariances are Estimated from Current Data
(n 1) 2
UCL Beta , p / 2,( n p 1) / 2 (10)
n
Generalized Variance Control Limits
UCL b1 3 b2 (11)
CL b1 (12)
LCL b1 3 b2 (13)
where
p
1
b1
(n 1) p
(n i )
i 1
(14)
1 p
p p
b2
(n 1) 2 p
(n i ) (n j 2) (n j ) (15)
i 1 j 1 j 1
If is not known, it is replaced by the estimate |S|/b1.
2009 by StatPoint Technologies, Inc. Multivariate EWMA Control Chart - 12
