Escolar Documentos
Profissional Documentos
Cultura Documentos
ACTIVE
MANIPULATION
OF VARIABLES
MAKES DOE A
BLACK BELTS
BEST BET.
By Christopher
Using statistical tools such as regression, a Six Sigma team can establish a correlation between variables of interest. But, correlation does not imply causation.
A small example illustrates this important, often overlooked, point.
Programmers at a large insurance firm were recently given the option of
attending a training workshop on object oriented programming, and about half
the eligible programmers volunteered. Each year the productivity of all programmers is rated on a 10-point scale.
Figure 1 compares the ratings of the programmers who attended the training
with those who did not. The supporting statistical analysis shows the observed difference in average productivity is highly statistically significant.
It is tempting to conclude the increase in productivity was caused by the training. But causality does not necessarily follow what was simply an observational
study of historical data. In this case, review of the prior years productivity scores
revealed the volunteers were already the most productive. So much for the effectiveness of the training.
So how do we establish causality? We must actively manipulate the variables we
Nachtsheim,
University of
Minnesota, and
Bradley Jones,
SAS Institute Inc.
9
8
7
6
5
no
yes
Volunteer
30
A U G U S T
2 0 0 3
W W W . A S Q . O R G
t-test
Assuming equal variances
Difference t-test DF Probability > ItI
Estimate
-1.2200 -4.776 28
<.0001
Standard error
0.2555
Lower 95%
-1.7433
Upper 95%
-0.6967
A P o w e r f u l A n a l y t i c a l To o l
C
C
N
C
Columns (4/0)
Temperature
Pressure
Supplier
Registration
error
Rows
All rows
Selected
Excluded
Hidden
Labeled
14
0
0
0
0
1
2
3
4
5
6
7
8
9
10
11
12
13
14
Registration
error
-15.4
-18.1
-18.8
-6.6
-15.8
4.6
-17.1
-10.6
3.7
3.1
10.4
-13.9
6.1
13.6
S I G M A
M A G A Z I N E
A U G U S T
2 0 0 3
31
A P o w e r f u l A n a l y t i c a l To o l
13.6
3.7
Supplier
pricey
cheap
25
30
20
300
325
-22.443
275
Registration error
13.6
-14.4111
32
A U G U S T
2 0 0 3
Supplier
W W W . A S Q . O R G
pricey
cheap
25
30
20
300
325
-22.443
275
Registration error
A P o w e r f u l A n a l y t i c a l To o l
17.8908
0.002355
-22.507
1
1
Supplier
pricey
cheap
29.0298
30
20
291.515
325
0
275
Desirablility
Registration error
Desirabililty
17.8908
0.00914
-22.507
1
0.999998
S I X
S I G M A
Supplier
pricey
cheap
30
30
275
291.406
20
325
Registration error
Desirablility
Figure 5 shows that the mean registration error at the current process settings
for the low cost supplier is about 14.4. The
scrap rate for this process is well over 50%,
which is clearly not a capable process.
By contrast, Figures 6 and 7 show the
optimal settings of temperature and pressure for each supplier, respectively. Figure
6 reveals lowering the temperature from
its current setting of 300 to 291 while raising the pressure from 25 to 29 psi moves
the process mean for registration error to
virtually zero. The scrap rate is 0.05%.
The big surprise is in Figure 7. Here,
lowering the temperature to 281 and raising the pressure to 30 psi with the low
cost epoxy leads to a zero predicted registration error and a scrap rate of 0.05%.
In the institutionalization step, the
improvement team recommended switching to the low cost supplier and standardizing the optimal process settings. The $5
million a year savings resulted from lower
material costs and reduced scrap.
Desirabililty
NOTE
REFERENCES
1. W. Edwards Deming, Out of the Crisis, MIT Press, 2000.
2. Mikel Harry and Richard Schroeder, Six Sigma, Currency, 2000.
3. Gray McQuarrie, Control of Key Registration Variables for Improved
Process Yields on Dense MLBs, IPC Expo 99, March 1999, Long Beach, CA.
4. Gray McQuarrie, Using DOE to Solve Compensation Problems, PC
Fab, April 2001.
5. JMP is statistical software produced by SAS Institute Inc.
M A G A Z I N E
A U G U S T
2 0 0 3
33