STAT 453/558: The Design and Analysis of Experiments

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

Chapter 7- Blocking and Confounding

Introduction
• In real experiments, nuisance factors (such as batch effects, time effects, or environ-
mental conditions) can introduce variation that is not of primary interest.
• Blocking allows us to control for this variation by ensuring that different treatment
combinations are compared within homogeneous groups (blocks), reducing unexplained
error.
• It is not always possible to perform all the runs in a 2k factorial experiment under
homogeneous conditions.
• Example: batch is not large enough to make all the required runs.
• We might want to run the experiment with several batches of material to account for
different quality grades.
• We can use blocking.

Why and how to block?

• If a 2k factorial design has been replicated n times, each replication is a block (and the
model is the same as discussed in Chapter 5).
• Each replicate is run in one of the blocks (time periods, batches of raw material, etc.)
• Within blocks, the runs are randomized.
• Blocking increases the power of the tests.

1
Blocking example in a 22 design

2
Confounding in the 2k factorial design
• Confounding is a technique that allows us to run fewer replicates while still estimating
the most important effects by grouping runs into blocks.
• When we confound higher-order interactions with blocks, we sacrifice estimation of
those interactions in exchange for blocking out unwanted sources of variation.
• Rationale: High-order interactions (e.g., three-way or four-way interactions) are often
assumed to be negligible in practice. We can deliberately confound these higher-order
interactions with the blocks to free up degrees of freedom for estimating lower-order
(main and two-factor interaction) effects.

When to confound?
• When there is a known source of variation that you cannot eliminate but can control
by grouping into blocks.
• When resource constraints prevent you from running a full factorial design without
blocking.
• When higher-order interactions are unlikely to be important, making it reasonable to
sacrifice them in favor of a more controlled experimental setup.

Example: Confounding the 22 factorial design in two blocks

• Consider a single replicate of a 22 design
• Each treatment combination requires raw material
• Each batch of raw material is only large enough for two treatment combinations
• Batch=block, and we must assign two treatment combinations to each block

How to confound?

• If we want to confound a treatment with blocks, we can assign all the positive signs
of the treatment to one block, and all the negative signs of the treatment to another
block
• Usually we don’t want main effects to be confounded with blocks
• The usual practice is to confound the highest order interaction with blocks

3
Methods for confounding
• Consider a 23 design: A, B, C
• 2 blocks, ABC confounded with blocks
• Methods
– Method 1: Use the tables of + and - signs
– Method 2: Use defining contrasts

Method 1

Method 2 - Defining Contrasts

• Based on the 0 and 1 notation (dummy variable)

• Works by using the linear combination

L = α 1 x1 + α 2 x2 + . . . + α k xk

• xi =0 for low level and xi = 1 for high level

• αi = 0 or αi = 1 (True for all 2k designs)

4
 • To assign the blocks we look at L(mod2), where mod2 is the modulo operation and
indicates the remainder when dividing by 2

For the 23 design

• For the treatment combination (1) we write as (0, 0, 0) in the (0, 1) notation

L = 1(0) + 1(0) + 1(0) = 0 = 0(mod2)

• a : L = 1(1) + 1(0) + 1(0) = 1 = 1(mod2)

• b : L = 1(0) + 1(1) + 1(0) = 1 = 1(mod2)
• ab : L = 1(1) + 1(1) + 1(0) = 2 = 0(mod2)
• c : L = 1(0) + 1(0) + 1(1) = 1 = 1(mod2)
• ac : L = 1(1) + 1(0) + 1(1) = 2 = 0(mod2)
• bc : L = 1(0) + 1(1) + 1(1) = 2 = 0(mod2)
• abc : L = 1(1) + 1(1) + 1(1) = 3 = 1(mod2)

• Thus (1), ab, ac, and bc are run in block 1 and a, b, c, and abc are run in block 2
• The block containing (1) is called the principal block

Estimation of the Error

• When k is small (say 2 or 3), it is usually necessary to replicate the experiment to
obtain an estimate of the error.
• If k is moderately large, say k ≥ 4, we can frequently afford only a single replicate.
• In the later cases, we assume higher order interactions to be negligible and combine
their sum of squares as errors.

5
Back to Example 6.2
• Factors are studied in a power plant to determine their effect on filtration rate
• 4 factors: A (temperature), B (pressure), C (formaldehyde), and D (stirring rate)
• Modifications to the original experiment:
– First modification: the 24 = 16 treatment combinations cannot all be run using
one batch of material (logical to confound ABCD with blocks).
– Second modification: introduce a block effect by subtracting 20 from the filtration
rate of runs in the principal block.

Table of contrasts for the 24 design

Example 6.2 in blocks

6
 • The previous interaction effect was 0.6875 × 2 = 1.375
• The introduced block effect is −20
• The block effect is 1.375 + (−20) = −18.675
• We can compute the block effect by
BE = ȳBlock1 − ȳBlock2 = 406/8 − 555/8 = −18.625
• This effect really estimates Blocks + ABCD.
• The sum of squares for blocks (Equation 4.11) is
b
1 X y2
SSBlocks = Bj − ...
nj j=1 N
where Bj is the sum of all observations in Block j, nj is the number of observations
in block j and y... is the sum of all observations. Note that the number of dots in the
subscript of y... changes depending on the number of factors in the model
(406)2 + (555)2 (961)2
SSBlocks = − = 1387.5625
8 16

Why blocking?
• Blocking is a very useful and important design technique.
• It has a dramatic potential to reduce the noise in an experiment.
• In doubt, block!

Example: What happens if we don’t block?

• Consider another variation of Example 6.2
• Previously, we constructed the design in two blocks, 8 runs each. We inserted a “block
effect” of magnitude −20 that affects all observations in block 1.
• Suppose that instead of having blocks we introduce the -20 nuisance factor impacting
the first 8 observations of the run.
• A new normal probability plot for this modified version shows us what happens.

• In this particular example, if we don’t block one of the important interactions AD is

not identified.
• If we don’t block, the added variability from the nuisance variable effect ends up getting
distributed across the other factors.
• Some of the variability ends up in the error estimate.
• In the example, the MSE is several times higher than in the original data.
• As we have seen before, blocking is a noise reduction technique.

7
Confounding the 2k in 4 blocks
• It is possible to construct 2k factorial designs confounded in four blocks of 2k−2 obser-
vations each.
• Particularly useful when the number of factors is large.

• If each block will hold only 4 runs, then we need 4 blocks

• Select 2 effects to be confounded with blocks
• As an example, let’s consider two effects ABC and ACD, with corresponding defining
contrasts
L 1 = x 1 + x2 + x3
L 2 = x1 + x3 + x4

• Every treatment combination will yield a particular pair of values L1 (mod2) and
L2 (mod2), that is (L1 , L2 ) = (0, 0), (0, 1), (1, 0), or (1, 1).
• Treatment combinations yielding the same values of (L1 , L2 ) are assigned to the same
block.

For this design we have 16 runs and we need to distribute them in four blocks.
• For the treatment combination (1) we write as (0, 0, 0, 0) in the (0, 1) notation

8
 • a : L1 = 1 + 0 + 0 = 1 = 1(mod2) and L2 = 1 + 0 + 0 = 1 = 1(mod2). Therefore
a = (1, 1).
• b : L1 = 0 + 1 + 0 = 1 = 1(mod2) and L2 = 0 + 0 + 0 = 0 = 0(mod2). Therefore
b = (1, 0)

• If we use the plus and minus table

ABC ACD Block AD

- - 1 +
- + 2 -
+ - 3 -
+ + 4 +

• Note that block has 3 degrees of freedom, but ABC and BCD have only 2 degrees of
freedom
• Add the generalized interaction of ABC and ACD given by ABC×ACD = A2 BC 2 D =
BD

General rules for confounding in 4 blocks

1. Choose two effects to generate the blocks, automatically confounding a third effect that
is the generalized interaction of the first 2.
2. When selecting effects to be confounded with blocks, be careful to not confound effects
of interest.
3. For example in a 25 design if we confound ABCDE and ABD, we automatically confound
CE, an effect that is probably of interest.
4. A better choice is to confound ADE and BCE, which automatically confounds ABCD.
5. It’s preferable to sacrifice information on the three-factor interactions instead of of two
factor interactions.

Confounding the 2k in 2p blocks

1. The methods described can be extended to the 2k design confounded in 2p blocks
(p < k).
2. Each block contains exactly 2k−p runs.
3. We select p independent effects to be confounded.
4. Independent means that no effect chosen is the generalized interaction of the others
5. Blocks can be generated by using the p defining contrasts L1 , L2 , . . . , Lp associated with
these contrasts.

9
6. In addition, 2p − p − 1 other effects will be confounded with blocks (the generalized
interactions of those p independent contrasts.

1. The statistical analysis of these designs is straightforward: sum of squares for all the
effects are computed as if no blocking had occurred.
2. The block sum of squares is found by adding the sum of squares for all the effects
confounded with blocks.
3. Table 7.9 shows a list of useful designs.

10