ESGC6112 – Lecture 7

Confounding/Blocking in 2k Designs (Pg 289)

7.1 Simple Confounding

(a) An unreplicated 2k experiment has 2k treatment combinations. If

k=3, we have 8 treatment combinations, and if each takes 2 hrs to
run, 16 hrs will be required to complete the experiment. We
cannot run all the 8 treatment combination as a large block under
homogenous condition, instead, we have to split it into two smaller
block of 4. We run this experiment over 2 days, a block of 4 on
Monday and another block on Tuesday. Many influences could
occur in the experiment, and make the interpretation of the results
unclear. For example: personnel could change etc.
(b) The 8 treatment combinations, in Yates‟ standard order are 1, a, b,
ab, c, ac, bc, abc; there are 70 ways we can assign them to the 2
days. Does it matter which of these to choose? Assigned randomly?
(c) Table 7.1 shows 3 of the 70 possible arrangements.
(d) Assume that if there is a „block effect‟, (the systematic difference
between a measurement made on Monday and the same one made
on Tuesday), it causes all Monday yields to be higher, by some
unknown but approximately constant amount, X, than if they had
been obtained on Tuesday; X can be positive or negative.
(e) Block that includes the „1‟ observation (all factors at low level) is
called the principal block.

Table 7.1: Three Possible Designs

Design 1 Design 2 Design 3
Monday Tuesday Monday Tuesday Monday Tuesday
Treatments 1 c 1 a 1 a
a ac ab b ab b
b bc c ac ac c
ab abc abc bc bc abc

(f) Design 1: all treatments with C high occur on Tuesday; all with C
low are on Monday. The C effect is estimated by taking the
difference between the average of the yields with C high and the
average with C low. With this blocking design, we can‟t tell how
much this difference is associated with the main effect C and how
1
 much is associated with the block (day) effect (and of course,
„noise‟ is present also). Estimate of C has thus been confounded
with the block effect, and other effects could be confounded as
well.
(g) Confounded – describe a situation in which 2 or more effects (C &
X) are estimated together, in sum or difference, and are not
separable.
(h) Design 2: The main effects are unaffected by the block effect – the
plus & minus sign or, equivalently, the number of high and low
levels of the factor are equally balance over the 2 days. But AB
are all positive on Monday, and negative on Tuesday. Hence, AB
interaction term is confounded with the block effect.
(i) Design 3: The main & interaction effects appear to be unaffected
by the block effect (the signs of the effects are „balance‟).
Usual estimate of main effect A is given by:
A (1/ 4)[ 1 a b ab c ac bc abc]
For design 3: The usual estimate of A:
A (1/ 4)[ (1 X ) a b (ab x) c (ac X ) (bc X ) abc]
A (1/ 4)[ 1 a b ab c ac bc abc] = usual estimate of A

The usual estimate of ABC is

ABC (1/ 4)[ 1 a b ab c ac bc abc]
in Design 3, estimate of ABC is
ABC (1/ 4)[ (1 X ) a b (ab X ) c (ac X ) (bc X ) abc ]
= usual estimate – X
The estimate of 3-factor interaction effect, ABC, is polluted by
(confounded with) the block effect; it is equal to the difference of
two quantities – the „pure‟ estimate of ABC and the block effect, X.

In Design 3, only the estimate of ABC has confounded; all main

and 2-factor interactions are unaffected by the separation of the 8
treatment combinations into 2 smaller blocks of 4.
(j) Main effects are most important, 2-factor interactions are of next
importance, and 3-factor interactions are of least interest, in an
experiment with 3 factors. Thus, Design 3 is probably superior to
the others.
(k) Another design as shown in Table 7.2.

2
 Table 7.2: A fourth Design
Monday Tuesday
Treatments 1 ab
a c
b bc
ac abc

The confounded effects are B, C, AB, AC.

Usual Estimate
A ( 1 a b ab c ac bc abc) / 4
B ( 1 a b ab c ac bc abc ) / 4
AB (1 a b ab c ac bc abc) / 4
C ( 1 a b ab c ac bc abc ) / 4
AC (1 a b ab c ac bc abc) / 4
BC (1 a b ab c ac bc abc) / 4
ABC (1/ 4)[ 1 a b ab c ac bc abc]

Fourth Design estimate

A [ (1 X ) (a X ) (b X ) ab c (ac X ) bc abc) / 4
= usual estimate of A
B [ (1 X ) (a X ) (b X ) ab c (ac X ) bc abc) / 4
= usual estimate of B – ½ X
AB [(1 X ) (a X ) (b X ) ab c (ac X ) bc abc) / 4
= usual estimate of AB – ½ X
C [ (1 X ) (a X ) (b X ) ab c (ac X ) bc abc) / 4
= usual estimate of C – ½ X
AC [(1 X ) (a X ) (b X ) ab c (ac X ) bc abc) / 4
= usual estimate of AC + ½ X
BC [(1 X ) (a X ) (b X ) ab c (ac X ) bc abc) / 4
= usual estimate of BC
ABC (1/ 4)[ (1 X ) (a X ) (b X ) ab c (ac X ) bc abc]
= usual estimate of ABC

(l) Of the 70 arrangements possible, only 14 have just one effect

confounded, the rest have more than one effect confounded, in
most cases four of them.

3
 (m) We can choose which effect we wish to confound. Thus, randomly
assigning treatments to blocks is silly because it abrogates an
opportunity to design a superior experiment.
(n) We choose a design that confounds only one effect- the most
expendable one, i.e. the highest-order interaction effect.
(o) Summary: It‟s best to run all treatment combinations at the same
time, under the same conditions. When we can‟t, we partition the
original block into 2 equal-sized smaller blocks. By doing so, we
reduce experimental error (that portion due to the block effect) for
the effects we care about most at the expense of confounding one
effect. All „clean‟ estimates can be judged against reduced
variability, with the corresponding narrower confidence limits,
increased power, and the like; for these more important effects, it‟s
as though there were no block effect.
(p) If the block effect is not independent of the treatment combinations,
but rather varies for different treatment combinations, then the
blocking factor (day or time) must be considered explicitly as
another factor under study.
(q) As the size of an experiment grows, confounding becomes more
popular due to it is difficult to create large homogeneous blocks,
and the loss of one effect may not be of great consequence.

7.2 Partial Confounding

23 Design with 4 replications

(a) Experiment being run over a four-week period, each replicate

consists of one block of 4 treatment combinations run on Monday,
the other block on Tuesday.
(b) The same design could be repeated each week, thereby
confounding the same effect, likely ABC. Each main effect and
each 2-factor interaction effect could be estimated based on all 32
data values.
(c) Each effect could be estimated by averaging the four estimates of
each effect, one per replicate, or equiv. averaging each of the 4
yields for the 8 treatment combinations before calculating the
effects.
(d) We could run a different “confounding scheme” or “design” for
each replicate, as shown in Table 7.3. This design „partially‟
confounds each of the interaction effects.
4
 Table 7.3: Four Replicates with Different Designs
Week 1 Week 2 Week 3 Week 4
Mon Tues Mon Tues Mon Tues Mon Tues
Treatment 1 a 1 a 1 a 1 b
ab b ab b b ab a ab
ac c c ac ac c bc c
bc abc abc bc abc bc abc ac
Confounded ABC AB AC BC
Effect

(e) The estimates of each of the three main effects are base on all 32
data values.
(f) We use all but the second replicate for AB, all but the third for AC,
all but the fourth for BC, and all but the first for ABC. The
interaction effects are each based on 24 data values. No effect is
completely confounded (or „lost‟); instead, each of the interaction
effects is partially confounded.

Example: Grading X-Ray Film

Dependent variable: film readability

Film readability is evaluated by experience radiologist using a reference
tumor pattern on a simulated torso.
Grading: 100-point scale; the higher the score, the greater the readability
Factors: (A) photo-resist thickness, (B) exposure time, (C) development time.
2 radiologists evaluate the films; each will examine four treatment
combinations for each of two replications.
The block effect is associated with the use of 2 different radiologists.
Two Partially Confounded Effects in 23Design
First Replication Second Replication
Dr. McGwire Dr. Sosa Dr. McGwire Dr. Sosa
Treatments 1 A 1 a
ab B ab b
ac C c ac
bc Abc abc bc
Coufounded ABC AB
Effect

5
A, B, C. AC, BC ARE clean in both replications; they will be estimated using
all 16 yields.
ABC and AB are partially confounded. ABC is estimated using data from 2nd
replication, and AB is estimated with data from 1st replication; both are
based on 8 data points.

7.3 Multiple Confounding

(1) As experiments grow larger, it is necessary to have > 2 blocks.

(2) Consider an unreplicated 24 design, partition the experiment into 4
blocks of 4 treatment combinations each. Table 7.4 shows the
possible design:
Table 7.4: Four Blocks in a 24 Design
Block 1 Block 2 Block 3 Block 4
1 a b c
cd acd bcd d
abd bd ad abcd
abc bc ac ab
Yields in the four blocks differ by constants – yields in the first block are
too high (or low) by R, those in the second block by S, those in the third
by T, and those in the fourth by U.

With no loss of generality, R + S + T + U = 0.

Total no. of effects = 15 (24 – 1).

The block effects confound three estimates: AB, BCD and ACD.

The minimum no. of confounded effects is number of blocks – 1.

(3) Table 7.5 illustrates a confounded effect (ACD) and 3 clean effects
(A, C and D).

6
 Table 7.5: Signs for A, C, D and ACD Effects
A C D ACD
Treatment Sign Block Sign Block Sign Block Sign Block
of trt Effect of trt Effect of trt Effect of trt Effect
1 - -R - -R - -R - -R
a + +S - -S - -S + +S
b - -T - -T - -T - -T
ab + +U - -U - -U + +U
c - -U + +U - -U + +U
ac + +T + +T - -T - -T
bc - -S + +S - -S + +S
abc + +R + +R - -R - -R
d - -U - -U + +U + +U
ad + +T - -T + +T - -T
bd - -S - -S + +S + +S
abd + +R - -R + +R - -R
cd - -R + +R + +R - -R
acd + +S + +S + +S + +S
bcd - -T + +T + +T - -T
abcd + +U + +U + +U + +U

Estimate of ACD = (1/8)(8ACD -4R+4S-4T+4U)

= ACD – R/2 + S/2 – T/2 + U/2

(4) In practice, the experimenters begin by deciding which effects they

would be willing to sacrifice (confound), and select the design to
achieve that result. Presumably the 24 experiment above was
designed to confound one 2-factor interaction, AB, and two 3-
factor interactions, BCD and ACD.
(5) Why not sacrifice ABCD? Barnard proved that in a 4-block design,
3 effects (at minimum) must be confounded, only 2 of the 3
confounded effects can be freely chosen by the designer; the third
is defined (i.e. mandated) by the choice of the first two. Insisting
on confounding ABCD will result in the loss of either 2 two-factor
interactions, or one main effect.

7
7.4 Mod-2 Multiplication

(a) The mod-2 multiplication of A and B is A B = AB, A and A is A A

= A2=A0=1(exponents are from the binary filed; the only
exponents allowed are 0 and 1, any odd-integer exponent is equiv.
to exponent 1 and any even-integer exponent is equiv. to exponent
0), AB and CD is AB CD = ABCD
(b) Barnard‟s theorem: in a 24 in four blocks, whenever two effects are
specified to be confounded, the mod-2 multiplication of these two
effects is automatically the third effect confounded. For a
threesome effects, it doesn‟t matter which two are specified; the 3
form a „closed group‟ of sorts. Table 7.6 illustrates this.

Table 7.6: Close Group of Confounded Effects

st
1 Specified 2nd Specified Resultant (3rd)
Confounded Effect Confounded Effect Confounded Effect
AB BCD AB BCD=AB2CD=
AB0CD=ACD
AB ACD AB ACD=A2BCD=
A0BCD=BCD
BCD ACD BCD ACD=ABC2D2=
ABC0D0=AB

(c) We need to choose the selected effects to be confounded with care;

otherwise the resulting 3rd confounded effect may not be a
desirable one, rendering the set of 3 confounded effects an unwise
choice. Suppose in a 25 design we had first specified the
confounding of ABCDE, and then ABCD, we would then lose
ABCDE ABCD = E, a main effect.
(d) It would be better to confound ABD, ACE and BCDE so as not to
lose any main effects or any two-factor interactions. – always
pick a design that minimizes the importance of the most important
effect confounded.

7.5 Determining the Blocks

(a) Treatment combinations that go into each block are determined as

follows:
8
 (i) Treatment combinations that have an even number of letters
in common with each of the confounded effects go into
block – the principal block;
(ii) Treatment combinations to fill the other blocks are
determined by mod-2 multiplication of any treatment
combination not in the principal block times each of the
treatment combinations in the principal block.

(b) Example: consider a 25 experiment with 4 blocks. 3 effects must

be confounded. We specified 2 out of 3: ABD and ACE, and
thereby confound BCDE. All main effects and 2-factor interactions
are clean.

Principal block: all treatment combinations that have an even (inc.

zero) number of letters in common with all confounded effects.

1st: list all 32 treatment combinations in Yates‟ order: 1, a, b,

ab, …., abcde. The principal block will be:
1 abc bd acd abe ce ade bcde

(i) 1 has an even number of letters in common with ABD, ACE,

and BCDE – zero. We always begin with 1 in forming the
principal block.
(ii) Identified abc, which has 2 letters in common with ABD, A
and B; it also has two in common with ACE, and also with
BCDE. How to identify? Look at first confounded effect
listed, ABD, and arbitrary picked the A and B as two in
common; then check with the next confounded effect list,
ACE, it didn‟t have the even number of letters in common,
so we added c, yielding abc, which have an even number of
letters in common with ACE, as well as being „even‟ with
ABD. abc needs to be „innovated‟.
(iii) Again using educated trial and error, we identify bd, which
has two letters in common with ABD, zero with ACE (and
two with BCDE).
(iv) Any two members of the principal block, when mod-2
multiplied, always yield a member of the principal block.
This yields abc bd = acd.
(v) Then innovated one more treatment combination: abe.

9
(vi) Finally, used the „closed group‟ property to multiply abe
times the previous 3 treatment combinations listed, to
generate, respectively, ce, ade, and bcde.

The elements abc, bd and abe, the treatment combinations that

were innovated, are said to be „generator‟ elements, they are not
unique, we could have ended up with other sets of three, depending
on which ones happened to think of first.

Second block:

(i) it is formed from the first by picking any yet-unused

treatment combination and mod-2 multiplying it by each
element of the principal block, respectively. If we pick a,
we get:

Principal block 1 abc bd acd abe ce ade bcde

Multiply by a a bc abd cd be ace de abcde

(block 2)

(ii) Continue the process by selecting some still-unused

treatment combination, say b, and mod-2 multiplying it by
the elements of the principal block:

Principal block 1 abc bd acd abe ce ade bcde

Multiply by a a bc abd cd be ace de abcde

(block 2)
Multiply by b b ac d abcd ae bce abde cde
(block 3)

10
(iii) The final block is made up of those treatment combinations
still unassigned. It may be easier to use one more still-
unused treatment combination, say e, mod-2 multiplying it
by the elements of the principal block. The results:

Principal block 1 abc bd acd abe ce ade bcde

Multiply by a a bc abd cd be ace de abcde

(block 2)
Multiply by b b ac d abcd ae bce abde cde
(block 3)
Multiply by e e abce bde acde ab c ad bcd
(block 4)

(iv) We compute all effects as before – by either using Yates‟

algorithm, or the sign table. Table 7.7 illustrates the
confounded nature of ABD, ACE, and BCDE and the
unconfounded nature of two arbitrarily chosen clean effects,
AB and D. for the clean effects, block effects cancel out
within each block, for both AB and D, the table of signs has
four plus sign and four minus signs within each block. This
equality if required to cancel out the block effects. Within
block, for any given counfounded effect, the signs are all the
same; rather than cancel, the block effects accumulate.

11
 Table 7.7: Sign Table for Confounded and Clean Effects

Confounded Effects Clean Effects

Block (and constant) Treatment ABD ACE BCDE AB D
Block 1 1 - - + + -
abc - - + + -
(Too high or too low bd - - + - +
by R) acd - - + - +
abe - - + + -
ce - - + + -
ade - - + - +
bcde - - + - +
Block 2 a + + + - -
bc + + + - -
(Too high or too low abd + + + + +
by S) cd + + + + +
be + + + - -
ace + + + - -
de + + + + +
abcde + + + + +
Block 3 b + - - - -
ac + - - - -
(Too high or too low d + - - + +
by T) abcd + - - + +
ae + - - - -
bce + - - - -
abde + - - + +
cde + - - + +
Block 4 e - + - + -
abce - + - + -
(Too high or too low bde - + - - +
by U) acde - + - - +
ab - + - + -
c - + - + -
ad - + - - +
bcd - + - - +

12
7.6 Number of Blocks and Confounded Effects

(1) Table 7.8 shows the minimum number of effects that will be
confounded, how many of these the designer can specify and how
many are an automatic consequence of the designer‟s choice.

Table 7.8: Number of Blocks and Confounded Effects

Number of Number of Number of Number Defined

Smaller Blocks Confounded Designer May by Consequence
Effects Choose
2r 2r-1 r 2r-1-r
2 1 1 0
4 3 2 1
8 7 3 4
16 15 4 11

(2) For the case of 8 blocks, the experimenter may initially choose 3 of
the necessary 7 effects to be confounded, X, Y and Z. Then, the 4
consequentially confounded effects will be:
X Y, X Z, Y Z, X Y Z

(3) For the case of 16 blocks, the experimenter may initially choose 4
of the 15 necessary effects to be confounded, X, Y, Z and V. Then
the 11 consequentially confounded effects will be these 4 effects
multiplied two at a time, three at a time, and four at a time:

X Y, X Z, X V, Y Z, Y V, Z V, X Y Z, X Y V, X Z V
Y Z V, and X Y Z V

Exercises :

Pg 308-309 : Questions 1, 2, 3, 7, 8, 15.

13