Issue 94, December 2008

Using DOE++ to Analyze Unbalanced Designs

[Editor's Note: This article has been updated since its original publication to reflect a more recent version of the software interface.]

One of the purposes of using Design of Experiments (DOE) is to evaluate which factors have significant effects and how they affect the response of interest to the experimenters. For example, one might examine how the speed and the type of gasoline affect the mileage of a vehicle and which factor has the greater impact. When DOE is used by reliability engineers, the response of interest is product life; a reliability engineer will use DOE to investigate which stresses influence product life and how much they affect life. In order to obtain an accurate estimation, experiments need to be carefully designed. A general requirement is that the design needs to be "balanced." In this article, we will discuss what an unbalanced design is and how it can be analyzed correctly using ReliaSoft's DOE++ software.

Definition: If all the treatments (combinations of factors) in an experiment have the same number of observations, the design is called a balanced design.

Instead of explaining the benefits of using balanced designs in a convoluted mathematical way, an intuitive and simple explanation is given here. For instance, in order to compare the hardness of the steel provided by two manufacturers, it is better to take the same number of samples from each manufacturer and compare their mean values. If you take only 1 sample from Vendor A and take 2000 samples from Vendor B, the resulting comparison will not be reliable, especially if the single sample from Vendor A is an outlier.

Analysis of Variance (ANOVA) is the technique that is usually employed for determining which factors are the major contributors to the variations of the response. In other words, this type of analysis is used to identity which factors have significant influences on the response. Once these factors are identified, actions can be taken by adjusting the factors to optimize the response. A typical ANOVA table for a two-level full factorial design using two factors looks like the one shown next.

Table 1: A Typical ANOVA Table

However, the method given in many DOE textbooks to calculate the sum of squares in the ANOVA table is applicable only for balanced designs.[1, 2] (Dr. S.R. Searle provides an excellent discussion of balanced designs.[3])

Although heavily unbalanced designs should be avoided, there are situations where it may be necessary to work with an unbalanced design. Many engineers may not be aware that the method taught in textbooks cannot be used to obtain correct results in such cases. Fortunately, many DOE software packages can handle unbalanced designs correctly. In this article, we will illustrate why the calculation method for balanced designs cannot be used for unbalanced designs and will provide the results calculated by DOE++.

Example 1: Balanced Design
This simple example is based on the following data:

Table 2: Data for a Balanced Design

It is a balanced design because it has an equal number of observations for each treatment (cell). There are only two observations for each treatment. The data can be entered into DOE++ using a two-level full factorial design with two factors and two replicates.

Figure 1: Example 1, Data Input in DOE++

For a balanced design, the sums of squares for factors A, B and AB can be calculated using the method given in the textbooks:[1, 2]

where:

• a is the number of levels for factor A.
• b is the number of levels for factor B.
• n is the number of observations of each cell or treatment.
• , , and are the corresponding, column, row, cell and grand average.

Expressed mathematically:

Using the above equations:

which matches the following results from DOE++ using the Individual Test Terms option:

Figure 2: Example 1, ANOVA Table in DOE++

The ANOVA table shows that factor A and the interaction effect AB are significant at the risk level of 0.1, while factor B is not. In other words, the p values for factor A and the interaction effect AB are less than the risk level used in the analysis of 0.10. Some readers may notice that the calculated sum of squares is "Partial Sum of Squares." In fact, there are several different types of sums of squares. DOE++ provides example files that offer discussions of many advanced topics in Design of Experiments, such as balanced vs. unbalanced design, partial vs. sequential sum of squares, etc. For additional details, please refer to the examples shipped with DOE++ (accessible by choosing Help > Open Examples Folder) and the accompanying reference book.[4]

Example 2: Unbalanced Design
If we add one more observation to the data in Example 1, it becomes an unbalanced design. The modified data set, which includes one more observation in cell 1 (A = Level 1, B = Level 1), is given in Table 3:

Table 3: Data for an Unbalanced Design

For this unbalanced design, the equations used in example 1 to calculate the sum of squares are not applicable. For example, if you calculate SSAB using an analogous equation:

Clearly, this result is wrong; the sum of squares cannot be negative, since squares cannot be negative. This result indicates that there is something wrong with the calculation method. Therefore, the method for the sum of squares for balanced designs cannot be used for unbalanced designs. Using DOE++, you will get the sum of squares as:

Figure 4: Example 2, ANOVA Table in DOE++ for Example 2

Instead of using the above formulas to calculate the sum of squares for each term, DOE++ always uses the linear regression method; using this method, SSAB is calculated to be 0.0606. From the results in Examples 1 and 2, we can see that the linear regression method can calculate the sum of squares correctly for both balanced and unbalanced designs, so it makes sense for DOE++ to use this method in all cases. For the details of the calculation, please refer to ReliaSoft's Experiment Design and Analyis Reference.[4]

Reference:
[1] Montgomery, D., Design and Analysis of Experiments, 5th edition, 2001, New York: John Wiley & Sons, 2001, p. 180.
[2] Wu,  C. F. and Hamada, M., Experiments: Planning, Analysis, and Parameter Design Optimization, New York: John Wiley & Sons, 2000, pp. 57-58.
[3] Searle, S. R. "A Biometrics Invited Paper: Topics in Variance Component Estimation," Biometrics, vol. 27, no. 1, pp 1-76.
[4] ReliaSoft Corporation, Experiment Design and Analysis Reference, Tucson, AZ: ReliaSoft Publishing, 2008.

Copyright 2014 ReliaSoft Corporation, ALL RIGHTS RESERVED