You are on page 1of 22

Lecture 13.

The 2k factorial design

Jesper Rydén
Matematiska institutionen, Uppsala universitet
jesper@math.uu.se

Regression and Analysis of Variance • autumn 2015


Example of 22 factorial design

Two factors A and B, each run at two levels (“low” and “high”).
Completely randomized experiment — the order in which the runs
are made is random.

A B Treatment combination
(1) − − A low, B low
(a) + − A high, B low
(b) − + A low, B high
(ab) + + A high, B high
Example of 22 design
Example: Chemical process, response variable is yield.
Factor A: reactant concentration (15 or 25 percent).
Factor B: amount of catalyst (1 pound or 2 pounds).
The experiment is replicated n = 3 times, in all 12 runs.
Example: R code and ANOVA table

y = c(28,36,18,31,25,32,19,30,27,32,23,29)
Conc = rep(c(-1,1),6)
Am = rep(c(-1,-1,1,1),3)
mod22 = lm(y ~ Conc + Am + Conc*Am)
anova(mod22)

Analysis of Variance Table

Response: y
Df Sum Sq Mean Sq F value Pr(>F)
Conc 1 208.333 208.333 53.1915 8.444e-05 ***
Am 1 75.000 75.000 19.1489 0.002362 **
Conc:Am 1 8.333 8.333 2.1277 0.182776
Residuals 8 31.333 3.917
---
Signif. codes: 0 *** 0.001 ** 0.01 * 0.05 . 0.1 1
R output, fitted model

summary(mod22)
Call:
lm(formula = y ~ Conc + Am + Conc * Am)

Residuals:
Min 1Q Median 3Q Max
-2.000 -1.333 -0.500 1.083 3.000

Coefficients:
Estimate Std. Error t value Pr(>|t|)
(Intercept) 27.5000 0.5713 48.135 3.84e-11 ***
Conc 4.1667 0.5713 7.293 8.44e-05 ***
Am -2.5000 0.5713 -4.376 0.00236 **
Conc:Am 0.8333 0.5713 1.459 0.18278
---
Signif. codes: 0 *** 0.001 ** 0.01 * 0.05 . 0.1 1
Matrices in a 22 design
Regression model for a general two-factor design:
y = β0 + β1 x1 + β2 x2 + β12 x1 x2
Let x1 denote factor A, x1 = −1 if low level, x1 = 1 otherwise.
Let x2 denote factor B, x2 = −1 if low level, x2 = 1 otherwise.
The matrix XT X is given as
 2 
2 n 0 0 0
2
0 2 n 0 0
XT X = 
 

 0 0 22 n 0 
0 0 0 22 n
where n is the number of observations per treatment combination.
Its inverse:
1/22 n
 
0 0 0
 0 1/22 n 0 0 
(XT X)−1 = 
 0 2

0 1/2 n 0 
0 0 0 1/22 n
Matrices in a 22 design

Matrix calculations lead to


 
(1) + a + b + ab
 a + ab − (1) − b 
XT y =  
 b + ab − (1) − a 
(1) + ab − a − b

and hence the LS estimate


((1) + a + b + ab)/22 n
 
2
b =  (a + ab − (1) − b)/2 n
 
β  (b + ab − (1) − a)/22 n


((1) + ab − a − b)/22 n

Further analysis and examples on blackboard!


Note on estimation of σ 2 for factorial experiment
Consider again a 22 design as a regression model:

y = β0 + β1 x1 + β2 x2 + β12 x1 x2 .

With n = 1, we have four runs — but also four parameters to


estimate! No degrees of freedom left for estimation of σ 2 .
Consider e.g. the test statistic for main effect A,

[a + ab − (1) − b]2 /(22 n)


FA = ∼ F (1, 22 n − 3 − 1)
SSE /(22 n − 3 − 1)
which with n = 1 becomes F (1, 0) distributed . . . (not defined!)
Possible solutions:
I Use σ 2 from previous experiments.

I Drop higher-order terms in the model and use these degrees of


freedom to estimate σ 2 (Montgomery: sparsity-of-effects
principle)
I Increase if possible n
The unreplicated factorial

A single replicate of a 2k design is sometimes called an


unreplicated factorial.
Montgomery, Section 6.6:

Unreplicated 2k designs are widely used in practice. They


may be the most common variation of the 2k design.
Yates’ standard order

Algebraic signs for calculating effects in the 23 design.


Properties of the table

Some interesting properties:


1. Except for the identity column I , each column has an equal
number of plus and minus signs.
2. Any two columns are orthogonal
3. Multiplying any column by column I leaves the column
unchanged. That is, I is an identity element.
4. The product of any two columns yields a column in the table,
e.g.
A × B = AB,
and
AB × ABC = A2 B 2 C = C
since any column multiplied by itself is the identity column.
The 23 design: Geometrical presentations

Geometric presentation of contrasts corresponding to main effects and


interaction in the 23 design. (a) Main effects. (b) Two-factor
interactions. (c) Three-factor interactions.
Example. Plasma Etch, a 24 design

Objective: Develop a nitride etch process on a single-wafer plasma


etcher.
Experiment with Etch rate as response and four factors Gap,
Pressure, C2 F6 flow and Power. Factor levels as follows:

A single replicate of a 24 design is available.

Solid State Technology (1987)


Example. Plasma Etch, a 24 design
Data from the 16 runs.
Example. Plasma Etch, a 24 design
Table of plus and minus signs.
Example. Plasma Etch, a 24 design
ANOVA table for the 24 design.
Main effects A, B, C , D significant, and interaction AD.
Example. Plasma Etch, a 24 design
Normal probability plot of the effects.
Example. Plasma Etch, a 24 design

Interaction plot, AD (Gap and Power).


Example. Plasma Etch, a 24 design
Normal probability plot of residuals. Montgomery and Runger:
“satisfactory”.
Example. Factorial experiments in two labs

The aim is to investigate the effect of two factors A and B on the


yield of a biochemical process. Each of two labs perform a 22
experiment. Assume that the observations from the labs have the
same variance σ 2 , but that a systematic difference could exist
between labs. Results:

Lab. 1 A B Yield Lab. 2 A B Yield


− − 32 − − 37
+ − 35 + − 42
− + 15 − + 21
+ + 27 + + 34
Example, cont.

For Lab. 1, the mean of the four observations is 27.25, for Lab. 2,
the corresponding mean is 33.50. If the eight observations are
regarded as one sample, its standard deviation is 8.879.
(a) Formulate a statistical model.
(b) Estimate main effects and interaction effects. Which effects
are statistically significant?
Example. An abbreviated experiment
Consider a 23 factorial experiment with the factors A, B and C
with response variable y (lifetime). However, there were
experimental difficulties when making the runs; the total
experiment was halted after only 4 runs.
Treatment Response
combination
a 43
b 35
c 44
abc 39

(a) Write down the table with signs for contrasts for the
performed experiment.
(b) Comment on the contrasts in your table. Are the contrasts
orthogonal? Are main effects orthogonal to each other?
This abbreviated experiment is called a fractional factorial.

You might also like