How To: Apply SPC Techniques to Correlated Variables

Using STATGRAPHICS Centurion

by

Dr. Neil W. Polhemus

July 27, 2005

Introduction
When analyzing the quality of manufactured products, it is common to measure more than one
variable. For example, two important characteristics of a textile fiber are its tensile strength and
its diameter. In applying SPC techniques to samples of data from such a process, it is common to
consider each variable separately, even though the variables are likely to be correlated. This can
lead to:
1. Erroneous estimates of the percent of the product that is beyond the specification limits.
2. Missed out-of-control signals on control charts.
In such cases, it is important to consider the variables simultaneously. In this guide, we will
examine two useful procedures in STATGRAPHICS Centurion: the Multivariate Capability
Analysis procedure and the Multivariate T-Squared Control Chart.

Sample Data
As an example, we will consider data on n = 200 fiber specimens. Measurements have been
made of two variables: tensile strength and diameter. A portion of the data is shown below:
Specimen
Tensile Strength Diameter
1
125.11
1.111
2
124.94
1.091
3
124.79
1.097
4
124.99
1.105
5
125.09
1.103
6
124.86
1.088
7
125.17
1.114
8
125.18
1.115
Figure 1: First 8 Rows of Sample Data

LSL
124
1.05

Target
125
1.10

USL
126
1.15

The specification limits for the fiber are:

124 tensile strength 126
1.05 diameter 1.15
The specification limits have been entered into additional columns in the data file.
2005 by StatPoint, Inc.

Step 1: Plot the Data

The first step in analyzing any new set of data is to plot it. With two variables, a simple X-Y Plot
will suffice:
Procedure: X-Y Scatterplot
To create an X-Y plot in STATGRAPHICS Centurion, simply press the X-Y Scatterplot
button on the main toolbar. On the data input dialog box, indicate the name of the columns
containing the data:

Figure 2: Data Input Dialog Box for X-Y Scatterplot

The resulting plot displays each of the 200 observations in the dataset:

Plot of Diameter vs Tensile Strength

1.13

Diameter

1.12
1.11
1.1
1.09
1.08
1.07
124

124.3 124.6 124.9 125.2 125.5 125.8

Tensile Strength
Figure 3: X-Y Scatterplot of Fiber Data
Notice the strong positive correlation between tensile strength and diameter.
2005 by StatPoint, Inc.

To help determine whether the relationship between the variables is linear, press the
Smooth/Rotate button
LOWESS smoother:

on the analysis toolbar. When the dialog box appears, request a Robust

Figure 4: Dialog Box for Scatterplot Smoothing

LOWESS stands for Locally Weighted Scatterplot Smoothing and is a technique that can be
applied to any X-Y scatterplot to help visualize the relationship between the variables plotted on
each axis. It is made robust or resistant to outliers by smoothing the data twice, downweighting points far from the first smooth when the second smooth is performed. The result is
shown below:

Plot of Diameter vs Tensile Strength

1.13

Diameter

1.12
1.11
1.1
1.09
1.08
1.07
124

124.3 124.6 124.9 125.2 125.5 125.8

Tensile Strength
Figure 5: X-Y Plot with LOWESS Smooth
The relationship does appear to be nearly linear.

2005 by StatPoint, Inc.

Step 2: Create Control Charts

To determine whether the process from which the data were obtained is in a state of statistical
control, control charts may be generated.
Procedure: Individuals Charts
Assuming that the samples were taken one at a time over some sampling period, the standard
approach is to construct an X chart and an MR(2) chart for each variable. In STATGRAPHICS
Centurion, this is done by:

If using the Classic menu, select: SPC Control Charts Basic Variables Charts Individuals.
If using the Six Sigma menu, select: Control Variables Control Charts Basic Control
Charts Individuals.

Complete the data input dialog box as shown below:

Figure 6: Data Input Dialog Box for Individuals Charts

The procedure creates an X-Chart and an MR(2) moving range chart. The X-Chart is shown
below:

2005 by StatPoint, Inc.

X Chart for Tensile Strength

125.8
UCL = 125.69
CTR = 124.96
LCL = 124.22

125.5

125.2
124.9
124.6
124.3
124
0

40

80

120

160

200

Observation
Figure 7: X-Chart for Tensile Strength
There are no observations beyond the 3-sigma control limits, although there are some runs rules
violations that seem to correspond to groups of similar values.
To plot Diameter, press the Input dialog button
variable:

on the analysis toolbar and change the data

Figure 8: Modified Data Input Dialog Box

The resulting X-Chart again shows no observations beyond the 3-sigma limits:

2005 by StatPoint, Inc.

X Chart for Diameter

1.14
UCL = 1.134
CTR = 1.099
LCL = 1.063

1.12
1.1
1.08
1.06
0

40

80

120

160

200

Observation
Figure 9: X Chart for Diameter
Again, there are no points beyond the 3-sigma control limits.
It should be noted that the control charts have been created in Initial Studies or Phase I mode,
allowing the data to determine the location of the centerline and control limits. As such, the
absence of points beyond the control limits leads to the conclusion that neither variable deviated
unexpectedly far from the average of the 200 samples.
Procedure: Multivariate Control Charts
Plotting separate control charts for each variable is common practice and will detect cases in
which a single variable is far from its expected value. When the variables are correlated,
however, this does not tell the whole story. An additional chart should also be created which
considers the joint properties of the variables. Such a chart would show any unusual samples
with respect to the variables when considered together.
STATGRAPHICS Centurion provides two control charts for plotting multivariate data: a chart
based on Hotellings T-Squared statistic, and a multivariate EWMA chart. To construct the Tsquared chart:

If using the Classic menu, select: SPC Control Charts Multivariate Control Charts
T-Squared Chart.
If using the Six Sigma menu, select: Control Variables Control Charts Multivariate
Control Charts T-Squared Chart.

Complete the data input dialog box as shown below:

2005 by StatPoint, Inc.

Figure 10: Multivariate T-Squared Control Chart Dialog Box

To create a Phase I chart, we need only enter the names of the variables to be considered. The
procedure will then calculate the means, variances, and covariances for the set of variables and
display the following chart:

Multivariate Control Chart

UCL = 11.54

T-Squared

24
20
16
12
8
4
0
0

50

100

150

200

Observation
Figure 11: Multivariate T-Squared Control Chart
A point is plotted for each of the n = 200 samples, with an upper control limit. For the fiber data,
one sample is way beyond the upper limit. Clicking on the red point with the mouse indicates
that this sample is specimen #183.
2005 by StatPoint, Inc.
7

A second very informative plot can be created by pressing the Graphs button
Control Ellipse:

and selecting

Control Ellipse
1.15

Diameter

1.13
1.11
1.09
1.07
1.05
124

124.3 124.6 124.9 125.2 125.5 125.8

Tensile Strength
Figure 12: 3-Sigma Control Ellipse
This plot shows the measurements for each of the 200 specimens, together with a control ellipse.
The control ellipse is similar to a 3-sigma limit, in the sense that it bounds 99.73% of the
expected distribution of the samples. It is based on a multivariate normal distribution, with its
centroid located at the sample means and the length and orientation of its axes calculated from
the sample covariance matrix.
Notice that the unusual specimen is well within the overall limits for both Tensile Strength and
Diameter. However, it is unusually weak for a specimen with its diameter. This sample would
only be identified as unusual by considering both variables together. An attempt to determine the
cause of the unusual specimen might well provide important information about the process.

Step 3: Calculate Process Capability

Except for the one unusual specimen, the process from which the textile specimens were
obtained appears to be fairly stable. It thus makes sense to perform a capability analysis. The
goal of such an analysis is to estimate the percentage of specimens that meet (or do not meet) the
specification limits for both tensile strength and diameter.
Procedure: Process Capability Analysis
If the variables were not correlated, a separate capability analysis could be done on each
variable. To perform a capability analysis on a single variable:

If using the Classic menu, select: SPC Capability Analysis Variables - Individuals.
If using the Six Sigma menu, select: Analyze Variable Data Capability Analysis
Individuals.

Beginning with Tensile Strength, the dialog input dialog box should be completed as shown
below:
2005 by StatPoint, Inc.

Figure 13: Data Input Dialog Box for Process Capability Analysis
The procedure creates a plot displaying a histogram of the sample data, a fitted normal
distribution, and several common capability indices:

Process Capability for Tensile Strength

frequency

40

LSL = 124.0, Nominal = 125.0, USL = 126.0

30

Normal
Mean=124.957
Std. Dev.=0.256557
Cp = 1.36
Pp = 1.30
Cpk = 1.30
Ppk = 1.24
K = -0.04

20
10
0
124

124.4

124.8

125.2

125.6

126

Tensile Strength
Figure 14: Process Capability Plot for Tensile Strength
Also displayed is a table of capability indices:

2005 by StatPoint, Inc.

Capability Indices for Tensile Strength

Specifications
USL = 126.0
Nom = 125.0
LSL = 124.0
Short-Term
Capability
0.24542
1.35821
4.24842
3.90086
3.90086
1.30029

Long-Term
Performance
Sigma
0.256557
Cp/Pp
1.29926
Zusl
4.06401
Zlsl
3.73153
Zmin
3.73153
Cpk/Ppk
1.24384
K
-0.04265
DPM
58.7163
119.321
Sigma Quality Level
5.35143
5.17416
Based on 6 sigma limits. Short-term sigma estimated from average moving range. The Sigma Quality Level
includes a 1.5 sigma drift in the mean.

Figure 15: Estimated Capability Indices for Tensile Strength

Of major interest are:
1. The one-sided capability indices Cpk and Ppk. These compare the distance from the
sample mean to the nearer specification limit, divided by 3 . Indices are shown using a
short-term estimate of sigma obtained from the moving range, and a long-term
estimate based on the sample standard deviation. The higher the value of these indices,
the better. Most companies have a goal of at least 1.33 for both indices.
2. The long-term estimate leads to an estimated defects per million (DPM) of 119.3. This
represents the estimated number of specimens in every million that would fall outside the
specification limits.
3. The Sigma Quality Level (SQL) is 5.17. In a typical Six Sigma program, the goal is to
achieve an SQL of 6, which equates to 3.4 DPM.
Repeating the same analysis on Diameter yields:

Process Capability for Diameter

LSL = 1.05, Nominal = 1.1, USL = 1.15

30

Normal
Mean=1.09859
Std. Dev.=0.0120239

frequency

25
20

Cp = 1.39
Pp = 1.39
Cpk = 1.36
Ppk = 1.35
K = -0.03

15
10
5
0
1.05

1.075

1.1

1.125

1.15

Diameter
Figure 16: Capability Plot for Diameter
2005 by StatPoint, Inc.

10

The calculated capability indices are:

Capability Indices for Diameter
Specifications
USL = 1.15
Nom = 1.1
LSL = 1.05
Short-Term
Capability
0.011948
1.39493
4.3028
4.06678
4.06678
1.35559

Long-Term
Performance
Sigma
0.0120239
Cp/Pp
1.38613
Zusl
4.27564
Zlsl
4.04111
Zmin
4.04111
Cpk/Ppk
1.34704
K
-0.0282
DPM
32.285
36.1486
Sigma Quality Level
5.49546
5.4686
Based on 6 sigma limits. Short-term sigma estimated from average moving range. The Sigma Quality Level includes a 1.5
sigma drift in the mean.

Figure 17: Capability Plot for Diameter

The long-term estimate of DPM for diameter is 36.1.
Procedure: Multivariate Capability Analysis
We have now estimated that about 119 of every million fibers will be out of spec with respect to
tensile strength and about 36 of every million out of spec with respect to diameter. If the
variables were independent, the total DPM would be very nearly the sum of the two, i.e., DPM =
155. Although the chance of being out of spec on both variables at the same time exists, it is very
small if the variables are independent:
Prob(out of spec on both if independent) =

119
36
0.004

=
1,000,000 1,000,000 1,000,000

To calculate the Sigma Quality Level of the combined variables, we can use the
STATGRAPHICS Centurion Six Sigma Calculator, found under Tools Six Sigma Calculator.
Enter the combined DPM in the top section and press the Calculate button:

2005 by StatPoint, Inc.

11

Figure 18: Six Sigma Calculator

The calculator shows that, considering the combined DPM for the two variables, the process is
operating at an SQL of approximately 5.11.
In fact, tensile strength and diameter are not independent, so it is incorrect to sum the DPM of
the two variables. Instead, the Multivariate Capability Analysis procedure should be used:

If using the Classic menu, select: SPC Capability Analysis Variables Multivariate
Capability.
If using the Six Sigma menu, select: Analyze Variable Data Capability Analysis
Multivariate Capability.

Enter the names of the variables in the study, together with the names of columns containing the
specification limits and target values:

2005 by StatPoint, Inc.

12

Figure 19: Data Input Dialog Box for Multivariate Capability Analysis
The latter columns contain one entry for each variable, as shown earlier in Figure 1. The Analysis
Summary produces a table that summarizes the observed and estimated DPM:
Multivariate Capability Analysis
Data variables:
Tensile Strength
Diameter
Number of complete cases: 200

Variable
Tensile Strength
Diameter

Sample
Mean
124.957
1.09859

Sample
Std. Dev.
0.256557
0.0120239

Variable
Tensile Strength
Diameter
Joint

Observed
Beyond Spec.
0.0%
0.0%
0.0%

LSL
124.0
1.05

Estimated
Beyond Spec.
0.0119321%
0.00361486%
0.0149248%

Nominal
125.0
1.1

USL
126.0
1.15

Estimated
DPM
119.321
36.1486
149.248

The StatAdvisor
This procedure determines the percentage of items beyond a set of multivariate specification limits. In this case, the estimated
frequency of non-conformities with respect to at least one of the 2 variables equals 149.248 per million.

Figure 20: Analysis Summary for Multivariate Capability Analysis

2005 by StatPoint, Inc.

13

Considering both variables together, the estimated DPM is 149.2, not 155.5. This is because the
strong positive correlation between the variables creates a non-trivial chance that when one
variable is out of spec, the other variable will be also.
If you press the Tables button
displayed:

on the analysis toolbar, a table of Capability Indices may be

Capability Indices
Index
Estimate
MCP
1.21
MCR
82.95
DPM
149.248
Z
3.61661
SQL
5.11661
Based on 6 sigma limits. The Sigma Quality Level includes a 1.5 sigma drift in the mean.

Figure 21: Table of Multivariate Capability Indices

In addition to the DPM, the table shows:
1. MCP a multivariate capability index similar to Cpk. It is equal to the value of Cpk that
yields the observed DPM when considering only one variable.
2. SQL the Sigma Quality Level corresponding to the estimated DPM. A DPM of 149.2
equates to an SQL of approximately 5.12.
It will be noted that the estimated number of defects is about 4% lower when allowing for the
correlation between the variables. Depending on circumstances, the difference could be larger or
smaller than in this example. The more variables that are considered and the stronger the
correlations between them, the more important it becomes to use the multivariate approach.
Also created by default is a Multivariate Capability Plot:
Multivariate Capability Plot
DPM = 149.248

123.5 124 124.5 125

125.5 126 126.5

Tensile Strength

1.16
1.12
1.08
1.04

Diameter

Figure 22: Plot of Fitted Multivariate Normal Distribution

This plot shows the multivariate normal distribution from which the estimate of DPM was
obtained. The red area corresponds to values that are outside the specification limits.
Also of interest is the Capability Ellipse:
2005 by StatPoint, Inc.

14

99.73% Capability Ellipse

MCP =1.21

1.15

Diameter

1.13
1.11
1.09
1.07
1.05
124

124.4

124.8

125.2

125.6

126

Tensile Strength
Figure 23: Capability Ellipse
The ellipse contains 99.73% of the multivariate normal distribution. The green rectangle
corresponds to the specification limits. A highly capable process will contain the ellipse
comfortably within the rectangle.

Conclusion
When applying SPC techniques to multiple variables, it can be important to treat them as a
multivariate sample rather than as independent measurements. If a strong correlation exists
between the variables, a multivariate control chart will sometimes pick up unusual events that
separate control charts for each variable would miss. When estimating the fraction of product
outside the specification limits, use of a multivariate normal distribution gives a more accurate
estimate of the DPM than simply summing the DPMs for each of the variables.
Although the computations required to treat variables from a multivariate perspective are
somewhat involved and can be time-consuming, it is important to consider related variables
together. Taking a multivariate approach can show features that might otherwise be missed,
which can lead to important process improvements.

Note: The author welcomes comments about this guide. Please address your responses to
[email protected].

2005 by StatPoint, Inc.

15