STATGRAPHICS – Rev.

9/16/2013

Gage Linearity and Accuracy

Summary ......................................................................................................................................... 1
Linearity Plot .................................................................................................................................. 6
Analysis Summary .......................................................................................................................... 7
Analysis Options ............................................................................................................................. 8
Accuracy Plot .................................................................................................................................. 8
Lack-of-Fit Test .............................................................................................................................. 9
Variation Barchart ........................................................................................................................... 9
Calculations................................................................................................................................... 10

Summary
The Gage Linearity and Accuracy procedure is designed to estimate the accuracy of a
measurement system. In contrast to procedures such as the Average and Range Method and the
Anova Method, which estimate the variability or precision of a measurement system, this
procedure is concerned with how accurate a system is on average.

To quantify the above concepts, suppose that a measurement process when repeated on the same
part many times yields measurements that come from a population with mean  and standard
deviation . The bias of the measurement process is defined as the difference between the mean
of that population and the true value being measured, i.e.,

bias =  - true value (1)

A process is said to be “accurate” if the bias is small. In contrast, precision is directly related to
, with smaller values of  being characteristic of “precise” processes. A poor measurement
process could be accurate but not precise, or precise but not accurate.

On the other hand, if a measurement is known to have substantial bias, but that bias is consistent
across all items being measured, then it may be possible to compensate for that bias by adjusting
the measured values. Linearity is a term that refers to how much the bias changes throughout the
normal operating range of a gage or other measurement instrument. If the bias is consistent, then
the linearity will be small.

Sample StatFolio: gagelinearity.sgp

 2013 by StatPoint Technologies, Inc. Gage Linearity and Accuracy - 1

 STATGRAPHICS – Rev. 9/16/2013
Sample Data:
The file linearity.sgd contains data from a typical gage linearity study, taken from the third
edition of the Automotive Industry Action Group’s (AIAG) reference manual on Measurement
Systems Analysis, MSA (2002). A portion of the data in that file is shown below:

Part Reference Measurement

1 2 2.7
2 4 5.1
3 6 5.8
4 8 7.6
5 10 9.1
1 2 2.5
2 4 3.9
3 6 5.7
4 8 7.7
5 10 9.3

The file contains a total of n = 60 rows, consisting of m = 12 measurements by a single appraiser

on each of g = 5 parts. The 5 parts have known reference values ranging from 2.00 to 10.00. The
intent of the study is to determine how well the measurement system matches those reference
values on average.

Note: Data reprinted from the Measurement Systems Analysis (MSA) Manual with
permission of DaimlerChrysler, Ford and GM Supplier Quality Requirements Task
Force.

 2013 by StatPoint Technologies, Inc. Gage Linearity and Accuracy - 2

 STATGRAPHICS – Rev. 9/16/2013
Data Input
The first dialog box displayed by this procedure is used to indicate the structure of the data to be
analyzed.

Input: The datasheet may be organized into either of two formats:

 Data and Code Columns: indicates that the datasheet contains a single column holding all the
measurements. This is the type of data structure illustrated above.
 One Row for Each Part: indicates that the datasheet contains a single row for all
measurements on a specific part. An example of this data structure is shown below:

Part Reference Trial 1 Trial 2 Trial 3 Trial 4 Trial 5 Trial 6

1 2 2.7 2.5 2.4 2.5 2.7 2.4
2 4 5.1 3.9 4.2 5 3.8 3.8
3 6 5.8 5.7 5.9 5.9 6 6.1
4 8 7.6 7.7 7.8 7.7 7.8 7.7
5 10 9.1 9.3 9.5 9.3 9.4 9.4

The second dialog box displayed depends on the setting in the first dialog box.

Data and Code Columns

If you select Data and Code Columns on the first dialog box, the second dialog box requests
the name of the column containing the measurements, as well as columns containing the part
numbers, reference values, and an optional value for the process variation.

 2013 by StatPoint Technologies, Inc. Gage Linearity and Accuracy - 3

 STATGRAPHICS – Rev. 9/16/2013

 Parts: numeric or non-numeric column indicating the an identifier for the item corresponding
to the measurement in each row.

 Measurements: numeric column containing the measurements.

 Reference Values: the true values of the variable being measured for each of the parts.

 Process Variation: the range over which the variable varies due to normal process variation.
If not specified, the process variation is set to the range of the Reference Values.

 Study Header: optional header to be printed at the top of each output table.

 Select: subset selection.

 2013 by StatPoint Technologies, Inc. Gage Linearity and Accuracy - 4

 STATGRAPHICS – Rev. 9/16/2013
One Row for Each Part
If you select One Row for Each Part on the first dialog box, the second dialog box requests the
names of multiple columns containing the measurements, as well as columns containing the part
numbers, reference values, and an optional value for the process variation.

 Parts: numeric or non-numeric column indicating an identifier for the item corresponding to
the measurements in each row.

 Measurements: two or more numeric columns containing the measurements.

 Reference Values: the true values of the variable being measured for each of the parts.

 Process Variation: the range over which the variable varies due to normal process variation.
If not specified, the process variation is set to the range of the Reference Values.

 Study Header: optional header to be printed at the top of each output table.

 Select: subset selection.

 2013 by StatPoint Technologies, Inc. Gage Linearity and Accuracy - 5

 STATGRAPHICS – Rev. 9/16/2013

Linearity Plot
The initial step in analyzing data from this type of study is to calculate the differences between
the observed measurements and the reference values:

yij = measurementij – reference valueij (2)

where measurementij represents the i-th measurement made on the j-th part. The Linearity Plot
shows these differences plotted against the references values xij:

Linearity Plot for Measurement

Y = 0.736667 - 0.131667X
1.4
R-Sq.=71.432%
1

0.6
Bias

0.2

-0.2

-0.6

-1
0 2 4 6 8 10

Reference
Included on the plot are:

 Blue squares: yij, the differences between the original measurements and the
corresponding reference values.

 Green diamonds: the average bias y j at each reference value.

 Blue line: the least squares regression of the differences yij against the reference values
xij. The equation of this line is shown along the top of the plot.

 Red lines: confidence limits for the regression line, at the level of confidence specified
on the Analysis Options dialog box.

 Green line: horizontal line at bias equal to 0.

 R.Sq.: the R-squared statistic for the fitted model. This is a measure of how well the line
fits the observed data.

In the current example, the bias changes quite a lot as the reference value changes, being positive
at lower reference values and negative at higher reference values. If the measurement system has
no bias, the green horizontal line will usually lie within the red confidence bands.

 2013 by StatPoint Technologies, Inc. Gage Linearity and Accuracy - 6

 STATGRAPHICS – Rev. 9/16/2013

Analysis Summary
The Analysis Summary displays summary statistics for the estimated linearity and accuracy of
the measurement system.

Gage Linearity and Accuracy - Part

AIAG Example
Parts: Part
Measurements: Measurement
Reference values: Reference
Process variation: 6

Number of Parts: 60
Total number of measurements: 60
Number of measurements excluded: 0
Range of reference values: 2 - 10

Estimates
Estimate Percentage
Bias -0.0533333 0.889
Linearity 0.79 13.167

Bias model: 0.736667 - 0.131667x

Parameter Estimate Stnd. Error t Statistic P-Value
Intercept 0.736667 0.0725243 10.1575 0.0000
Slope -0.131667 0.0109334 -12.0426 0.0000

Important information includes:

 Bias: the estimated average difference between the measured values and the reference
values. The percentage bias is also calculated by:

% bias = |bias| / (process variation) (3)

and shows the magnitude of the bias relative to the normal operating range of the process.

 Linearity: the estimated change in the bias over the normal variation of the process. This
is calculated by:

linearity = |slope| (process variation) (4)

The percentage linearity is calculated by:

% linearity = linearity / (process variation) (5)

and shows how much the bias changes as a percentage of the process variation.

 Bias model: the equation of the fitted regression line and t-tests for the significance of
the coefficients. Of particular interest is the P-value for the slope. If this value is small,
(less than 0.05 if operating at the 5% significance level), then there is significant change
in the bias over the range of the reference values.

In the current example, the average bias is small (less than 1%), but it changes by more than 13%
over the operating range of the process. Notice also that the P-Value for the slope is well below
0.05. The process would therefore normally be classified as having a problem with linearity.
 2013 by StatPoint Technologies, Inc. Gage Linearity and Accuracy - 7
 STATGRAPHICS – Rev. 9/16/2013

Analysis Options

 Decimal Places for Percentages: number of decimal places to use when displaying
percentage bias and linearity.

 Confidence Level: level used when plotting confidence bands on linearity and accuracy
plots.

Accuracy Plot
The Accuracy Plot shows an estimate of the accuracy of the measurement process.

Accuracy Plot for Trial 1

Y = -0.736667 + 0.131667X
1.3
R-Sq.=71.432%
0.9
Accuracy

0.5

0.1

-0.3

-0.7

-1.1
0 2 4 6 8 10

Reference
Accuracy is defined as the difference between the true value and the measured value:

accuracy = reference value – measured value (6)

It is similar to the Linearity Plot with 2 major differences:

1. The Y-axis is inverted, since accuracy and bias are opposite in sign.
2. The bounds on the plot show the estimated prediction limits for individual observations,
not confidence limits for the line.

You can sometimes detect the presence of outliers by examining points outside the prediction
limits, such as the two very low points at a reference value of 4. If the data was entered in the
Data and Code Column format, then the Exclude button on the analysis toolbar can be used to
interactively remove any outliers and recalculate the regression line.
 2013 by StatPoint Technologies, Inc. Gage Linearity and Accuracy - 8
 STATGRAPHICS – Rev. 9/16/2013

Lack-of-Fit Test
This pane shows the result of a test performed to determine whether a linear model is adequate to
describe the change in bias with reference value.

Analysis of Variance with Lack-of-Fit

Source Sum of Squares Df Mean Square F-Ratio P-Value
Model 8.32133 1 8.32133 145.02 0.0000
Residual 3.328 58 0.0573793
Lack-of-Fit 0.188 3 0.0626667 1.10 0.3579
Pure Error 3.14 55 0.0570909
Total (Corr.) 11.6493 59

Of primary interest is the P-Value on the Lack-of-Fit line of the table. If the P-value is small (less
than 0.05 at the 5% significance level), then there is statistically significant “lack-of-fit”,
meaning that a linear model is not adequate to describe the observed relationship. Such a result
would imply the need for a curvilinear model. There is not significant lack-of-fit in the current
example, since the P-Value is 0.3579.

Variation Barchart
This pane displays a barchart showing the percentage linearity and the percentage bias.

Percent of Process Variation

15
13.167%
12
Percent

3
1.036%
0
Linearity Bias

 2013 by StatPoint Technologies, Inc. Gage Linearity and Accuracy - 9

 STATGRAPHICS – Rev. 9/16/2013

Calculations

 Number of Measurements at j-th reference value

nij (7)

 Average bias at j-th reference value

nj

y ij
yj  i 1
(8)
nj

 Bias
g
nj  yj
j 1
y g
(9)
n
j 1
j

 Regression
The bias model is calculated by regressing the individual differences yij on the reference
values xij. This has changed since Version 5 of STATGRAPHICS, where a weighted
regression was performed on the average biases at each reference value. The estimated model
will be different only if the study is unbalanced, i.e., if there are different numbers of
measurements for different parts. The R-Squared statistic, however, will typically be smaller
than in previous versions for all studies. This change was made to reflect changes in the latest
edition of the AIAG manual.

The details of the regression calculations, including the t-tests, confidence bands, and lack-
of-fit test, are given in the documentation for the Simple Regression procedure.

 2013 by StatPoint Technologies, Inc. Gage Linearity and Accuracy - 10