You are on page 1of 28

Historical DOE

Using pre-existing data and


regression for analyses

1
Module Objectives

By the end of this module, the participant will:


• Perform and analyze a “Historical DOE”
• Define potential applications in Black Belt projects
SM

2
Why Historical DOE?

• Easy to do so, “Why not do it?”


• Inexpensive
• Uses available data
• Used in all applications with continuous data
• Identifies potential sources of variation
• Helps Black Belt focus where to work
• Increases process knowledge (caution)

3
What is Historical DOE?

• Advanced regression application


• Passive analysis
• Used on existing, but potentially “NOISY” data
– Expect regression to have lower R-square
– Non-numerical input data must be changed to
numerical for regression

4
Historical DOE Method

1. Arrange data sets using rational subgroups


2. Check integrity of data
3. Define Custom Response Surface Design
4. Analyze Response Surface Design
5. Reduce the regression model
6. Check residuals
7. Create model
8. Practical interpretation

Y = f (X )

5
Minitab HISTORICAL DOE

GATHER HISTORICAL
DATA
ANALYZE AS YOU
WOULD A NORMAL
RS DESIGN &
REDUCE
STAT> DOE>
RESPONSE
SURFACE>DEFINE
CUSTOM RS DESIGN USE REDUCED
MODEL TO
PREDICT
OUTCOMES (NOISY
DEFINE FACTORS USE DATA)
CODED DEFAULT
SETTINGS

EXPECT LOWER
THAN NORMAL R-
SQUARES
STAT> DOE> ANALYZE
RS DESIGN> TERMS>
FULL QUADRATIC

6
Step 1. Gather Historical Data

• Determine output(s)
• Align selected input(s) with output(s)
• Data could come from other DOEs
– Range is limited to these input values
– Range of input(s) determine inference space

7
Step 2. Check Integrity of Data

A B C Y
80 55 72.5 1382
75 55 69.5 1420
80 55 66.5 1436
85 55 69.5 1427
80 60 69.5 1437
85 60 72.5 1415
85 60 66.5 1442
80 60 69.5 1461
75 60 72.5 1422
75 60 66.5 1467
80 60 69.5 1447
80 65 72.5 1444
85 65 69.5 1455
80 65 66.5 1460
75 65 69.5 1449

8
Preparation for Step 3

Tool Bar Menu > Stat> DOE > Response Surface > Define Custom Response Surface Design

• Open worksheet: historicaldata example.mtw


• Data sets in columns

9
Step 3. Define Custom Response
Surface Design

10
Step 3. Define Custom Response
Surface Design

11
Step 4. Analyze Response Surface
Design Minitab Commands
Tool Bar Menu > Stat> DOE > Response Surface > Analyze Response Surface Design

12
Step 4. Analyze Response Surface
Design Minitab Commands

Note: In Storage, save the fits and residuals on last reduction

13
Step 4. Analyze Response Surface
Design

Term Coef SE Coef T P


Constant 1448.33 6.580 220.116 0.000
A -2.37 4.029 -0.589 0.581
B 17.88 4.029 4.436 0.007
C -17.75 4.029 -4.405 0.007
A*A -2.29 5.931 -0.386 0.715
B*B -8.29 5.931 -1.398 0.221
C*C -9.54 5.931 -1.609 0.169
A*B -0.25 5.698 -0.044 0.967
A*C 4.50 5.698 0.790 0.465
B*C 9.50 5.698 1.667 0.156

S = 11.40 R-Sq = 90.4% R-Sq(adj) = 73.1%

14
Step 5. Reduce the Regression Model

Term Coef SE Coef T P


Constant 1442.29 3.854 374.218 0.000
B 17.87 3.605 4.958 0.001
C -17.75 3.605 -4.923 0.001
C*C -8.79 5.277 -1.665 0.127
B*C 9.50 5.099 1.863 0.092

S = 10.20 R-Sq = 84.6% R-Sq(adj) = 78.5%

15
Step 5. Reduce the Regression Model
(Continued)

Term Coef SE Coef T P


Constant 1437.60 2.837 506.740 0.000
B 17.87 3.885 4.601 0.001
C -17.75 3.885 -4.569 0.001
B*C 9.50 5.494 1.729 0.112

S = 10.99 R-Sq = 80.4% R-Sq(adj) = 75.0%

16
Step 6. Check Residuals Commands

Tool Bar Menu > Stat> Regression > Residual Plots

17
Step 6. Check residuals

Residual Model Diagnostics


Normal Plot of Residuals I Chart of Residuals
30 UCL=28.59
20
20
10
Residual

10

Residual
0 0 Mean=9.11E-14

-10
-10
-20
-20 -30 LCL=-28.59

-2 -1 0 1 2 0 5 10 15
Normal Score Observation Number

Histogram of Residuals Residuals vs. Fits


5
20
4
10
Frequency

Residual
3
0
2

1 -10

0 -20
-20 -15 -10 -5 0 5 10 15 20 25 140014101420143014401450146014701480
Residual Fit

18
Step 7. Graphical Interpretation
Commands

19
Step 7. Graphical Interpretation
Commands

20
Step 7. Graphical Interpretation

Contour Plot of Y
72.5 1420
1430
71.5 1440
1450
70.5 1460
1470
69.5
C

68.5

67.5

66.5
55 56 57 58 59 60 61 62 63 64 65
B

21
Step 7. Create model (continued)

CODED
Y = 1437 . 6 + 17 . 87 ( B )
- 17 . 75 ( C )

UNCODED

Y = 1634.31 + 3.57500 (B) - 5.91667 (C)

22
Step 8. Practical Interpretation

• Look at:
– Residual Plots
– Contour Plots
– Formulas
• Based on current process knowledge look for factors and
trends that point to potential KPIVs

23
Historical DOE Exercise

A Black Belt gathered average sales data from a


database.

The Output (Y) terms are Total Sales (K$/day)


The work team reduced the Probable “X” factors

The Input terms are: A) $ on advertising, B) Hours of


training and C) Scores on a Customer Satisfaction
survey

The goal is to achieve $18,000/day at the least cost.

24
Historical DOE Exercise (continued)

1. Divide into work groups


2. Open HISTORICAL Sales.MTW
3. Use Historical DOE Method
4. Create a report (Power Point)
5. Prepare a report for class
6. Total time 30 Minutes

25
Application Exercise

• Break into small groups


• Decide how you could use this method for your projects
• Be prepared to present to the class

26
Key Learning Points







27
Module Objectives

By the end of this module, the participant should:


• Perform and analyze a “Historical DOE”
• Define potential applications in Black Belt projects

28