Chapter Contents
Chapter Contents
Previous
Previous
Next
Next
Details of the FACTEX Procedure

Example 15.14: Design with Inner Array and Outer Array

See FACTEX4 in the SAS/QC Sample Library

Byrne and Taguchi (1986) report the use of a fractional factorial design to investigate fitting an elastomeric connector to a nylon tube as tightly as possible. Their experiment applies the design philosophy of Genichi Taguchi, which distinguishes between control factors and noise factors. Control factors are typically those that the engineer is able to set under real conditions, while noise factors vary uncontrollably in practice (though within a predictable range).

The experimental layout consists of two designs, one for the control factors and one for the noise factors. The design for the control factors is called the inner array, and the design for noise factors is called the outer array. The outer array is replicated for each of the runs in the inner array, and a performance measure ("signal-to-noise ratio") is computed over the replicate. The performance measure thus reflects variation due to changes in the noise factors. You can construct such a cross-product design with the replication options in the OUTPUT statement of the FACTEX procedure, as shown in this example.

Researchers identified the following four control factors that were thought to influence the amount of force required to pull the connector off the tube:

Researchers also identified the following three noise factors related to the assembly:

Three levels were selected for each of the control factors, and two levels were selected for each of the noise factors.

The following statements construct the 72-run design used by Byrne and Taguchi (1986). First, an eight-run outer array for the three noise factors is created and saved in the data set OUTERARY.

   proc factex;
      factors time temperat humidity;
      output out=outerary time     nvals=( 24 120)
                          temperat nvals=( 72 150)
                          humidity nvals=(.25 .75);
   run;

Next, a nine-run inner array (design of resolution 3) is chosen for the control factors. The POINTREP option in the OUTPUT statement replicates the eight-run outer array in the data set OUTERARY for each of the nine runs in the inner array and saves the final design containing 72 runs in the data set SAVEDESN.

   proc factex;
      factors interfer connwall idepth glue / nlev=3;
      size design=9;
      model resolution=3;
      output out=savedesn pointrep=outerary
           interfer cvals=('Low'     'Medium' 'High'  )
           connwall cvals=('Thin'    'Medium' 'Thick' )
           idepth   cvals=('Shallow' 'Deep'   'Medium')
           glue     cvals=('Low'     'High'   'Medium');
   run;

   proc print data=savedesn;
   run;

The final design is listed in Output 15.14.1. Main effects of each factor can be estimated free of each other but are confounded with two-factor interactions.

Output 15.14.1: Design for Control Factor and Noise Factors
 
Obs interfer connwall idepth glue time temperat humidity
1 Low Thin Shallow Low 24 72 0.25
2 Low Thin Shallow Low 24 72 0.75
3 Low Thin Shallow Low 24 150 0.25
4 Low Thin Shallow Low 24 150 0.75
5 Low Thin Shallow Low 120 72 0.25
6 Low Thin Shallow Low 120 72 0.75
7 Low Thin Shallow Low 120 150 0.25
8 Low Thin Shallow Low 120 150 0.75
9 Low Medium Medium Medium 24 72 0.25
10 Low Medium Medium Medium 24 72 0.75
11 Low Medium Medium Medium 24 150 0.25
12 Low Medium Medium Medium 24 150 0.75
13 Low Medium Medium Medium 120 72 0.25
14 Low Medium Medium Medium 120 72 0.75
15 Low Medium Medium Medium 120 150 0.25
16 Low Medium Medium Medium 120 150 0.75
17 Low Thick Deep High 24 72 0.25
18 Low Thick Deep High 24 72 0.75
19 Low Thick Deep High 24 150 0.25
20 Low Thick Deep High 24 150 0.75
21 Low Thick Deep High 120 72 0.25
22 Low Thick Deep High 120 72 0.75
23 Low Thick Deep High 120 150 0.25
24 Low Thick Deep High 120 150 0.75
25 Medium Thin Medium High 24 72 0.25
26 Medium Thin Medium High 24 72 0.75
27 Medium Thin Medium High 24 150 0.25
28 Medium Thin Medium High 24 150 0.75
29 Medium Thin Medium High 120 72 0.25
30 Medium Thin Medium High 120 72 0.75
31 Medium Thin Medium High 120 150 0.25
32 Medium Thin Medium High 120 150 0.75
33 Medium Medium Deep Low 24 72 0.25
34 Medium Medium Deep Low 24 72 0.75
35 Medium Medium Deep Low 24 150 0.25
36 Medium Medium Deep Low 24 150 0.75
37 Medium Medium Deep Low 120 72 0.25
38 Medium Medium Deep Low 120 72 0.75
39 Medium Medium Deep Low 120 150 0.25
40 Medium Medium Deep Low 120 150 0.75
41 Medium Thick Shallow Medium 24 72 0.25
42 Medium Thick Shallow Medium 24 72 0.75
43 Medium Thick Shallow Medium 24 150 0.25
44 Medium Thick Shallow Medium 24 150 0.75
45 Medium Thick Shallow Medium 120 72 0.25
46 Medium Thick Shallow Medium 120 72 0.75
47 Medium Thick Shallow Medium 120 150 0.25
48 Medium Thick Shallow Medium 120 150 0.75
49 High Thin Deep Medium 24 72 0.25
50 High Thin Deep Medium 24 72 0.75
51 High Thin Deep Medium 24 150 0.25
52 High Thin Deep Medium 24 150 0.75
53 High Thin Deep Medium 120 72 0.25
54 High Thin Deep Medium 120 72 0.75
55 High Thin Deep Medium 120 150 0.25
56 High Thin Deep Medium 120 150 0.75
57 High Medium Shallow High 24 72 0.25
58 High Medium Shallow High 24 72 0.75
59 High Medium Shallow High 24 150 0.25
60 High Medium Shallow High 24 150 0.75
61 High Medium Shallow High 120 72 0.25
62 High Medium Shallow High 120 72 0.75
63 High Medium Shallow High 120 150 0.25
64 High Medium Shallow High 120 150 0.75
65 High Thick Medium Low 24 72 0.25
66 High Thick Medium Low 24 72 0.75
67 High Thick Medium Low 24 150 0.25
68 High Thick Medium Low 24 150 0.75
69 High Thick Medium Low 120 72 0.25
70 High Thick Medium Low 120 72 0.75
71 High Thick Medium Low 120 150 0.25
72 High Thick Medium Low 120 150 0.75

Note that the levels of IDEPTH and GLUE are listed in the OUTPUT statement in a nonstandard order so that the design produced by the FACTEX procedure matches the design of Byrne and Taguchi (1986). The order of assignment of levels does not affect the properties of the resulting design. Furthermore, design can be randomized with the RANDOMIZE option in the OUTPUT statement.

Byrne and Taguchi (1986) indicate that a smaller outer array with only four runs would have been sufficient. You can generate this design (not shown here) by modifying the statements on Example 15.14; specifically, add the following SIZE and MODEL statements:

     size design=4;
     model resolution=3;

In their analysis of the data from the experiment based on the smaller design, Byrne and Taguchi (1986) note several interesting interactions between control and noise factors. However, since the inner array is of resolution 3, it is impossible to say whether or not there exist interesting interactions between the control factors. In other words, you cannot determine whether an effect is due to an interaction or to the main effect with which it is confounded. One alternative is to begin with a design of resolution 4. Two-factor interactions will remain confounded with one another, but they will be free of main effects. Moreover, further experimentation can be carried out to distinguish between confounded interactions that seem important. To determine the optimal size of this design, submit the following statements interactively:

   proc factex;
      factors interfer connwall idepth glue / nlev=3;
      model resolution=4;
      size design=minimum;
   run;

This causes the following message to appear in the SAS log:

NOTE: Design has 27 runs, resolution = 4.

In other words, the smallest resolution 4 design for four three-level factors has 27 runs, which together with the eight-run outer array requires 216 runs. Even the smaller four-run outer array requires 108 runs. Both of these designs are substantially larger than the design originally reported, but the larger designs protect against the effects of unsuspected interactions.

A second alternative is to begin with only two levels of the control factors. Further experimentation can then be directed toward exploring the effects of factors determined to be important in this initial stage of experimentation. Note that NLEV=2 is the default in the FACTORS statement. Submit the following additional statements:

      factors interfer connwall idepth glue;
      model resolution=4;
      size design=minimum;
   run;

This causes the following message to appear in the SAS log:

NOTE: Design has 8 runs, resolution = 4.

Thus, as few as eight runs can be used for the inner array. This design is amenable to blocking, whereas the proposed nine-run design is not. Blocking is an important consideration whenever experimental conditions can vary over the course of conducting the experiment.

Now, submit the following statements:

size design=8;
blocks size=minimum;
run;

This causes the following message to appear in the SAS log:

NOTE: Design has 8 runs in 4 blocks of size 2,
resolution = 4.

Thus the experiment can be run in blocks as small as two runs.

Chapter Contents
Chapter Contents
Previous
Previous
Next
Next
Top
Top

Copyright © 1999 by SAS Institute Inc., Cary, NC, USA. All rights reserved.