Example 3: Gage
Repeatability and Reproducibility
you are manufacturing small kilns that are used for drying materials used
in other manufacturing processes. Assume for this example that the temperature
range at which these kilns operate is usually between 90 and 110 degrees
Celsius (°C). Before
performing a process capability analysis (see Example
2: Process Capability Analysis), you want to ensure that the measurement
system you are using is sufficiently precise to detect variability between
kilns. A repeatability and reproducibility (R & R) study will first
be designed to assess the precision of the measurement system, and then
the results of that experiment will be analyzed.
Designing the R &
R experiments. For the first part of this example, it is irrelevant
which data set is open, but a data set must be opened to access the Process Analysis Procedures. Since
we will use Temperat.sta in the
second part of this example, open this data file and start Process
bar. Select the Home tab.
In the File group, click the
Open arrow and on the menu, select
Open Examples to display the
Open a STATISTICA Data File dialog
box. Open the Temperat.sta data
file, which is located in the Datasets
folder. Then, select the Statistics
tab. In the Industrial Statistics
group, click Process Analysis
to display the Process Analysis Procedures Startup
menus. On the File menu,
select Open Examples to display
the Open a STATISTICA Data File
dialog box. Open the Temperat.sta data
file, which is located in the Datasets
folder. Then, on the Statistics - Industrial
Statistics & Six Sigma submenu, select Process
Analysis to display the Process Analysis Procedures Startup Panel.
Double-click Gage repeatability &
reproducibility to display the Repeatability & Reproducibility Analysis
- Generate design dialog box.
Now, suppose you have five engineers who are routinely involved in the
production process. Those engineers will serve as your operators of the
gages; thus, enter 5 in the Number of operators box.
Also, assume that within the available time frame, you can manage to
run a study where each engineer measures 8 kilns (parts) three times (3
trials). Therefore, enter 8 as
the Number of parts and 3 as the Number
of trials in the respective boxes on the Generate design tab. The Repeatability & Reproducibility Analysis
dialog box - Generate
design tab will now look like this:
Note that you can also specify single-trial designs with the Process
Analysis module. In that case, the repeatability and reproducibility
of measurements cannot be separately estimated, but will both contribute
to the variability across operators measuring the same part. This type
of "abridged" R & R study is useful in order to perform
periodic checks on the precision of the measurement system.
Click the OK button in
the Repeatability & Reproducibility Analysis
- Generate design dialog box to display the Repeatability & Reproducibility Design
The design of the R & R study can be displayed (and saved) either
in the Standard STATISTICA data file
format (using the Data files tab), or the Standard
gage R & R data sheet format (using the R & R data sheets tab) without
any grouping or coding variables (see, Gage
Repeatability and Reproducibility Overview for details).
For this example, choose the Standard
STATISTICA data file format, i.e., use the Data
files tab. In this format, the design summary can be displayed
in the spreadsheet in two ways: in Randomize
trials or in Standard order,
which is selected by default. Leave the default option button selected,
and then click the Summary: Display
design button in the Standard
STATISTICA data file format group box to create the spreadsheet.
Note that a blank column (Measure)
is added to the spreadsheet so that when you print this spreadsheet, you
will have a convenient data entry form. Note also that you can save this
design in a standard STATISTICA
data file (so that you can later enter the measurements and analyze the
data, see below) via the Save design
Part of the Standard order
design summary spreadsheet is shown below:
When you select the Randomize trials
option button, STATISTICA will
randomize the parts within operators and trials. It is always recommended
to randomize the experiment in this manner in order to rule out any serial
For example, operators may tire, and measurements taken later in the
experiment may be less accurate than those taken early on. Part of the
Randomize trials design spreadsheet
is shown below.
Standard R & R data sheet. This
way of displaying the design in a spreadsheet is particularly suited if
you want to use the (printed) spreadsheet to facilitate data entry. Shown
below is the summary of the design displayed in this format. (Click the
Summary: Display design button
on the R & R data sheets tab.)
As you can see, in this format there are no grouping variables (columns)
in the spreadsheet; instead, each column represents the data for one part.
This format of presenting the R & R experiment is often used in the
applied literature (e.g., ASQC/AIAG,
Analyzing the R &
R Experiment. Now, assume that you have completed the study and
recorded the measurements in the data file Temperat.sta.
Close all the spreadsheets and/or workbook in STATISTICA
except the Temperat.sta data
Note that the file has been "enhanced" by adding in the first
variable of the file the last names of the engineers as alphanumeric values.
In this way you can more easily identify the operators later in the analysis.
Click the Cancel button in
the Repeatability & Reproducibility Design
dialog box to close it and return to the Repeatability & Reproducibility Analysis
- Generate design dialog box. Select the Analyze data file tab.
The operator names (codes) are given in variable OPERATOR,
the part numbers are given in variable PART,
the trial numbers (codes) in variable TRIAL,
and the measurements in variable MEASURE.
Click the Variables button to
display a standard variable
specification dialog box and then specify those variables, as shown
Click the OK button to return
to the Analyze data file tab. Then, click
the Codes: (for operators, parts, trials)
button, and in the resulting dialog box, select all codes by either clicking
the All button for each variable
or the Select All button to select
all codes for all of the variables.
Click the OK button to return
to the Analyze data file tab, which will
now look like this:
Finally, click the OK button
to proceed to the Gage Repeatability & Reproducibility Results
This dialog box contains options to review the results of the analysis.
To graphically view the results of this study, you can choose from several
types of plots. First, click the Repeatability
& reproducibility plot button on the Quick tab.
Interpreting the plot. The individual
points plotted here are the deviations of the respective measurements
from the average measurement for the respective part. For example, look
at the furthest upper-left point (an approximate value of 4.4).
This point represents a measurement of Part
1 by engineer Smith; specifically
it shows that the measurement deviated from the average measurement for
part 1 by about plus 4.4 degrees Celsius. The other two measurements made
by Smith on the same part are
connected to the point via the vertical line. Thus, you can see that the
two other measurements deviated from the average for Part
1 by about 3.0 and -0.4. The second vertical line (to
the right) represents the three measurements of Part
2 made by engineer Smith,
and so on. Here is how you can interpret the information in this plot:
Location of the box. Each operator
is represented by a box. The location of this box allows you to identify
operators who are generally biased in all three measurements.
For example, operator Hill
"sticks out," in that his measurements are generally below the
average measurements of the same parts by other engineers. The average
deviation of the respective operator's measurements is also indicated
by the dashed horizontal line in each box.
Height of the box. The height
of the box is an indication of the variability of measurements across
trials. In our example, it seems that operator Miller
generally produced less variability across measurements on the same part
than any of the other engineers.
Length of vertical lines connecting
points. These lines connect the different trials by each operator
for each part. Therefore, excessively long lines identify wide ranges
of measurements on the same part, and thus greater imprecision; if all
trials yield identical measurements, there are no lines at all, but only
The "perfect" plot. What
would this plot look like if you had perfect repeatability and reproducibility?
If there were perfect repeatability, all repeated measurements on the
same parts would be identical, and you would see no vertical lines. If
there were perfect reproducibility, all operators would produce identical
measurements, therefore, the location of all boxes with regard to the
y-axis would be identical at value 0.
Analysis of Variance (ANOVA) results.
Now look at some numerical output. As described in the Process
Analysis - Gage Repeatability and Reproducibility Overview and Technical
Notes, the Process Analysis
module can estimate the different variance
components either from ranges or from the ANOVA table. The former
method is still popular because most of the necessary computation can
be done by hand; however, the ANOVA method is more accurate and should
thus be preferred (see ASQC/AIAG,
1990, page 65).
After returning to the Results dialog box, select the Advanced tab. Click the Complete
ANOVA table button. Two spreadsheets will be produced; the first
one shows the sums of squares for all effects.
If you are not familiar with the ANOVA method or with the concept of
main effects and interactions,
it is recommended that you read the Introductory
Overview to the ANOVA/MANOVA
module, which discusses these concepts and provides examples.
It is customary in R & R studies to regard the variability due to
the interactions involving the Trials
factor as error variability. This assumption seems reasonable, since,
for example, it is difficult to imagine how the measurement of some parts
will be systematically different in successive trials, in particular when
parts and trials are randomized. The next spreadsheet shows the ANOVA
results, treating all interactions by Trials
Looking at the results in this spreadsheet it appears that the Operator by Parts interaction is not
statistically significant. Thus, we could ignore this interaction and
consider a simpler ANOVA model without this interaction (select the No 2-way (Operator-Part) interaction
check box on the Advanced tab).
Now review the right-most columns of the spreadsheet shown above. In
these columns you find the estimates for the variances
deviations) for the components of interest. Once again, consider what
you would like your ideal measurement system to look like. Ideally, if
you had a perfectly repeatable and reproducible measurement system, then
all operators would arrive at identical measurements regardless of trial.
Therefore, there would be no variability due to operators (perfect reproducibility),
no variability due to trials (perfect repeatability), but only variability
due to parts.
Components of variance. To express
the variance components in terms of percentages of the total variability,
click the ANOVA method variance estimate
button on the Advanced tab.
The last column of numbers reports the variability due to different
sources relative to the total variability in the measurements: Repeatability
of measurements accounts for 6.5%,
reproducibility across appraisers accounts for 8.1%
of the total variability, the part-to-part variation accounts for 85.4%, and the combined repeatability
and reproducibility variability accounts for about 14.6%
of the total process variability. Thus, most of the variability in measurements
is due to differences between parts, as is desirable for a reliable measurement
system. Using the common guidelines for evaluating the quality of the
measurement system (under 10% = OK, 10% to 30% = questionable, above 30%
= needs improvement; see ASQC/AIAG,
1991, page 127), these percentages indicate that the performance of the
measurement system is acceptable.
You could now proceed to use this measurement system to put a quality
control system (chart) in place (use the Quality Control Charts module), to evaluate
your machine capability, or to use designed experiments to improve the
quality of your process (use the Experimental Design module). Now, review
some additional results available from the Gage Repeatability & Reproducibility Results
Identifying outliers. The idea
behind (Shewhart) quality control charts is discussed in the Introductory
Overview to the Quality Control Charts module. In short,
you can produce a chart of a quality characteristic across different samples
and establish control limits to identify outliers.
In this case, you are particularly interested in identifying outliers
among your engineers, that is, to identify those operators who produce
particularly imprecise measurements; you may also want to identify outliers
among parts, that is, to identify those parts that are particularly difficult
to measure, and therefore yield imprecise measurements. This type of information
may help you locate the causes of imprecision in your measurement system.
To identify outliers with regard to measurement precision, you want
to chart the variability of measurements across trials. Two standard charts
for controlling the variability of a process are the R chart of ranges
and the S (sigma) charts of standard deviations;
both can be produced by operators or by parts.
On the Descriptives/plots tab, click the
Sigma chart by operator button
to produce the plot shown below.
Interpreting the S chart. What
you see charted here are the standard deviations for the three measurements
trials by each engineer (operators) on each of the 8 kilns (parts). The
average standard deviation is indicated by the solid lines is at 2.03. The plus 3 times sigma
limit for this chart is at 5.21;
this upper control limit is indicated by the dashed line. Again, if you
are not familiar with this common choice for an upper control limit, refer
to the Introductory
Overview of the Quality Control Charts module. To review,
would fall above that sigma limit,
and any such outliers indicate that the respective engineer produced particularly
unreliable measurements for the respective part. However, you are mostly
interested in detecting general patterns here.
For example, as you saw in the summary plot produced earlier, engineer
Miller seemed to have produced
the least variability across trials; perhaps by observing how engineer
Miller is using the temperature
gages, you can find out how to make your measurement system even more
Box and whisker plot. Now look
at one other plot: the box
and whisker plot. This plot provides another summary of the average
measurements across trials and across parts. It also allows you to examine
whether the measurements are normally
distributed, an assumption that must be met before most quality control
procedures can be applied (e.g., control charts, process capability analysis,
After returning to the Descriptives/plots tab, click the
Box & whisker plot button.
For each operator, this plot summarizes the range of average measurements
(averaged across trials) as well as the distribution of those average
measurements. In this case, for each operator the median
seems to fall in the upper part of each box. The median statistic and
ranges are discussed in greater detail in the Basic Statistics and Tables Introductory
Overview, and in the Nonparametrics Introductory
Overview. In short, each box denotes the range of values into which
the center 50 percent of all measurements fall. The median itself "splits"
the distribution in half; that is, it is the point below and above which
50% of all measurements fall. In a normal
distribution, the mean
is equal to the median and would fall in the center of each box. In our
plot, however, the distributions appear to have a "long tail"
towards the lower end of the measurements.
In this particular example, the quality control engineer might want
to look at the distributions of measurements more closely to determine
how serious is the deviation from the normal distribution. In general,
unless the distribution is clearly not normal, most procedures are not
seriously affected. If such cases should occur though, normality can usually
be achieved by applying appropriate transformations to the measurements.
For example, log transformations will "pull in" the lower tail
of the distribution, etc.