STAT 453/558: The Design and Analysis of Experiments

Michelle F. Miranda
University of Victoria
[email protected]
Office: DTB 543

Chapter 6 - The 2k Factorial Design

Introduction
The 2k factorial design is a special case of the general factorial design described on Chapter
5.

• There are k factors, each at only two levels

• The two levels are usually called low and high (they could be either quantitative or
qualitative)
• The design is most efficient when studying the effects of two or more factors

Example on Yield of a Chemical Process

• Investigation into the effect of the concentration of the reactant and the amount of
catalyst on the conversion (yield) of a chemical process
• Objective: determine if adjustments to either of these two factors would increase the
yield
• Factor A: reactant concentration, 2 levels: 15% and 25%
• Factor B: catalyst, 2 levels: 2 pounds and 1 pound
• Experiment is replicated 3 times, 12 runs
• The order of the runs is random, so this is a complete randomized experiment

1
Estimation of the Effects

• In a 22 design we define the average effect of a factor as the change in response

produced by a change in the level of that factor averaged over the levels of the
other factors.

Practical interpretation

• The effect of A is positive, suggesting that increasing A from the low level (15%) to
the high level (25%) will increase the yield
• The effect of B is negative, suggesting that increasing the amount of catalyst added to
the process will decrease the yield
• The interaction effect seems small relative to the two main effects

2
ANOVA for the 22 Design
• It’s convenient to write down the treatment combinations in the order (−1), a, b, ab
(standard order)

• Contrasts are used to estimate the sum of squares for A, B, and AB

ContrastA = ab + a − b − (1),
this contrast is called the total effect of A.
• The average effect of A is given by
(ContrastA ) ab + a − b − (1)
A= k−1
=
2 n 2n
• The sum of squares are given by
(Contrast)2
SS =
2k n
• Therefore

3
 [ab + a − b − (1)]2
SSA =
4n
[ab + b − a − (1)]2
SSB =
4n
[ab + (1) − a − b]2
SSAB =
4n
• To estimate the MS we need the degrees of freedom as bellow

SS DF
SSA 1
SSB 1
SSAB 1
SSE 4(n-1)
SST 4n-1

• It’s important to examine the magnitude and direction of the factor effects along with
ANOVA because ANOVA does not convey this information. This will tells us which
variables are likely to be important.

• The average effects are

1
A= [90 + 100 − 60 − 80] = 8.33
2×3
1
B= [90 + 60 − 100 − 80] = −5.00
2×3
1
AB = [90 + 80 − 100 − 60] = 1.67
2×3
• And the sum of squares are
[90 + 100 − 60 − 80]2 502
SSA = = = 208.33
4×3 12
[90 + 60 − 100 − 80]2
SSB = = 75.00
4×3
[90 + 80 − 100 − 60]2
SSAB = = 8.33
4×3
• The effect of A is positive suggesting that increasing the reactant concentration from
the low to the high level will increase the yield.
• The effect of B is negative suggesting that increasing the amount of catalyst will de-
crease the yield
• The interaction effect appears to be small relative to the main effects.

4
The regression model
• For the chemical process example

y = β0 + β1 x1 + β2 x2 + β12 x1 x2 

• Assuming that the interaction term is zero, the fitted model is

8.33 −5
ŷ = 27.5 + x1 + x2
2 2

• The intercept is the average of all 12 observations

• The regression coefficients are 1/2 of the effect estimate.

The 23 Design
• 3 factors: A, B, and C,
• Each factor at two levels low (-) and high (+)
• Similarly to the 22 , we can list the 8 runs in a geometric view
• We write the treatment combinations in standard order as (1), a, b, ab, c, ac, bc, and abc
• Remember these symbols also represent the total of all n observations taken at that
particular treatment combination

Contrasts
• Three different notations are used for the runs in the 2k design: geometric coding (+ or
-), lowercase letter labels, the use of 0 and 1 to denote the low and high level (dummy
variable)

• There are 7 degrees of freedom between the 8 treatment combinations in the 23 design
• 3 degrees of freedoms are associated with the main effects A, B, and C
• 4 degrees of freedom are associated with the interactions (one each): AB, AC, BC, and
ABC

5
6
Estimating the main effect A
• The effect of A when B and C are at the low level is [a − (1)]/n
• The effect of A when B is at the high level and C is at the low level is [ab − b]/n
• The effect of A when C is at the high level and B is at the low level is [ac − c]/n
• The effect of A when both B and C are at the high level is [abc − bc]/n

• The average effect of A is just the average of these 4

1
A= [a − (1) + ab − b + ac − c + abc − bc]
4n
• The effect of A is just the average of the four runs where A is at the high level (ȳA+ )
minus the average of the four runs where A is at the low level (ȳA− )
A = ȳA+ − ȳA−
a + ab + ac + abc (1) + b + c + bc
= −
4n 4n
1
= [a + ab + ac + abc − (1) − b − c − bc] (1)
4n

Estimating the main effect B and C

• Similarly the effect of B is
B = ȳB + − ȳB −
1
= [b + ab + bc + abc − (1) − b − c − ac] (2)
4n
• Similarly the effect of C is
C = ȳC + − ȳC −
1
= [c + ac + bc + abc − (1) − a − b − ab] (3)
4n

7
The two-factor interaction effects

• A measure of the AB interaction is the difference between the average A effects and
the two levels of B. By convention, 1/2 of this difference is called the AB interaction

• AB is one-half of the difference shown above.

abc − bc + ab − b − ac + c − a + (1)
AB = (4)
4n
abc + ab + c + (1) bc + b + ac + a
= −
4n 4n
• Similarly we can find

abc + ac + b + (1) bc + c + ab + a
AC = −
4n 4n
1
= [(1) − a + b − ab − c + ac − bc + abc] (5)
4n

1
BC = [(1) + a − b − ab − c − ac + bc + abc] (6)
4n

The three-factor interaction effect

• The ABC interaction is defined as the average difference between the AB interaction
at the two different levels of C

1
ABC = [abc − bc − ac + c − ab + b + a − (1)] (7)
4n

• The quantities in brackets on equations (1) to (7) are contrasts.

8
Algebraic Signs in the 23 design

• A table of plus and minus signs can be developed from the contrasts

Properties of the factorial effect table

• Each column has an equal number of plus and minus signs

• The sum of the products of the signs in any two columns is zero
• The product of any two columns yields a column in the table
– A × B = AB
– AB × B = A
– A × ABC = BC

Sum of Squares of the effects

• In the 23 design with n replicates, the effect is given by

Contrast
Effect =
4n

• In the 23 design with n replicates, the sum of squares for any effect is
(Contrast)2
SS =
8n

Significance of the effects

• We can use ANOVA to to determine which factors effects are nonzero

9
 • We can construct CI for the effects
CI = Effect ± s.e.(Effect)

 
Contrast
V (Effect) = V
n2k−1
• Each contrast is a linear combination of 2k treatment totals, and each consists of n
observations
V (Contrast) = n2k σ 2
1
V (Effect) = k−2 σ 2
n2
Therefore
q √
\ = √M SE ,
s.e.(Effect) = V (Effect)
2
n2k
since σ̂ 2 = M SE.

The general 2k design

• k factors each at 2 levels
• 2k runs for each replicate, n replicates
• Model:
– k main effects
– k2 two-factor interactions

– k3 three-factor interactions...

– kk k-factor interaction

Total of 2k − 1 effects
• Error has 2k (n − 1) degrees of freedom
Properties

ContrastAB = (a − 1)(b − 1)(c + 1) . . . (k + 1)

ContrastBC = (a + 1)(b − 1)(c − 1)(d + 1) . . . (k + 1)
• Effects:
2
AB = (ContrastAB )
n2k
• Sum of Squares
1
SSAB = (ContrastAB )2
n2k
• F-test
M SAB
F =
M SE

10
A single replicate of the 2k design
• n=1
• 2k observations
• Effects have 2k − 1 d.f. = total d.f.
• Error: 0 d.f. (No MSE)
• No F-tests to test for effects

Inference when n = 1
Since there is no degrees of freedom left for the MSE, we can use use the following strategy.
• The effects that are negligible are usually normally distributed with mean 0 and variance
σ2
• Significant effects will have nonzero means
• A normal probability plot can be used to identify important factors

From a normal probability plot we can identify significant effects and keep them in the re-
duced model. Non-significant effects will not be in the reduced model, and they are combined
as an estimate of the error.

Maximizing and minimizing y

• We can find which level combination maximizes or minimizes the response variable y
• Regression model is a very useful way to optimize y
• We look at ŷ to find optimum levels for the factors

11
Example 6.2

• In R we can obtain the estimated effects (Remember 2× the regression coefficient)

• We won’t be able to tell whether the coefficients are significant
• We can use a normal probability plot to determine significant effects
• To maximize (or minimize) y we can look at the fitted values ŷ

• By looking at the normal probability plot we noticed that only the main effects of A,
C, and D and the AC and AD interactions are important.
• We can discard B from the experiment so that the design becomes a 23 design in A, C,
and D with two replicates. This is called design projection.
• Note that by projecting the single replicate 24 into a replicated 23 design we now have
both an estimate of the ACD interaction and an estimate of the error based on what
is called the hidden replication.

12
Projecting 2k design into 2k−1
• If one factor is not important in a 2k design, we can project the 2k design into a 2k−1
design with 2 replicates
• Assume we have a 23 design and B is not a significant effect. Then, the
• Assume we have a becomes a 22 design with 2 replicates as shown below

A B C A C
- - - - -
+ - - + -
- + - - -
+ + - + -
- - + - +
+ - + + +
- + + - +
+ + + + +

• All effects in a 2k design are orthogonal

• Because of orthogonality, the coefficients of the reduced model are the same as the full
model
• Therefore no need to run the model again without the non-important factors

Model Adequacy Checking Solutions

Data transformation in a factorial design

• If we find clear problems with normality and equality of variance we can transform the
response variable y
• Common transformations are
– y ∗ = ln y
√
– y∗ = y + c
– y ∗ = (y λ − 1)/λ (called the Box-Cox transformation)

13