2^k Factorial Experiments and Confounding
CSIR NET factorial designs: 2^2 and 2^3 effects, Yates' algorithm, sums of squares, total and partial confounding, principal block and generalized interaction.
- 18 min
On this page · 13 sections
Prerequisites
- CRD, RBD and Latin Square Designs: a factorial experiment is usually laid out as an RBD; only the treatment SS is split further.
- Block Designs: Connectedness, Orthogonality and BIBD: confounding is a way of running a factorial in incomplete blocks.
- Groups, Subgroups and Homomorphisms: the \( 2^k \) treatment combinations form the group \( \mathbb Z_2^k \); the principal block is a subgroup and the other blocks are its cosets.
- The chapter map is on Sampling Theory and Design of Experiments.
The Hook: Eight Treatments, Blocks of Four
Notation for \( 2^k \) Factorials
Factors \( A, B, C, \dots \), each at a low and a high level. A treatment combination is written by the lower-case letters of factors at the high level: \( (1) \) (all low), \( a \), \( b \), \( ab \), \( c \), \( ac \), \( bc \), \( abc \), … The same symbols also denote the totals of the \( r \) replicates of that combination.
Standard (Yates) order: \( (1), a, b, ab, c, ac, bc, abc, d, ad, \dots \), each new letter multiplying everything before it.
The \( 2^2 \) experiment
The simple effect of \( A \) at low \( b \) is \( a - (1) \); at high \( b \) it is \( ab - b \). The main effect is their average, and the interaction is half their difference:
$$A = \frac{1}{2r}\left[ab + a - b - (1)\right] = \frac{1}{2r}(a - 1)(b + 1),$$$$B = \frac{1}{2r}\left[ab + b - a - (1)\right] = \frac{1}{2r}(a + 1)(b - 1),$$$$AB = \frac{1}{2r}\left[ab - a - b + (1)\right] = \frac{1}{2r}(a - 1)(b - 1).$$The products are expanded symbolically and \( (1) \) is read for the constant term.
General rule
For \( 2^k \) with \( r \) replicates, the effect \( X \) (a main effect or interaction, written as a set of capital letters) is
$$X = \frac{1}{r\,2^{k-1}}\,[X], \qquad [X] = \prod_{\text{factors}}(\ell \pm 1),$$with the sign \( - \) for letters in \( X \) and \( + \) for letters not in \( X \). For example, in a \( 2^3 \), \( [AC] = (a - 1)(b + 1)(c - 1) \).
| Effect | (1) | a | b | ab | c | ac | bc | abc |
|---|---|---|---|---|---|---|---|---|
| A | − | + | − | + | − | + | − | + |
| B | − | − | + | + | − | − | + | + |
| AB | + | − | − | + | + | − | − | + |
| C | − | − | − | − | + | + | + | + |
| AC | + | − | + | − | − | + | − | + |
| BC | + | + | − | − | − | − | + | + |
| ABC | − | + | + | − | + | − | − | + |
Sign rule. For a main effect, \( + \) where its letter is present. For an interaction, multiply the signs of its letters. Equivalently, the sign of \( X \) at a treatment combination \( t \) is \( (-1)^{\lvert X \rvert - \lvert X \cap t \rvert} \).
Sums of squares and variances
The \( 2^k - 1 \) contrasts \( [X] \) are mutually orthogonal (the sign columns are orthogonal), and each has coefficient sum of squares \( 2^k \) on totals of \( r \) plots. Hence
$$\boxed{SS_X = \frac{[X]^2}{r\,2^k}, \qquad V(\hat X) = \frac{\sigma^2}{r\,2^{k-2}},}$$each SS on 1 df, and \( \sum_X SS_X = \) treatment SS. In an RBD with \( r \) replicates: blocks \( r - 1 \) df, treatments \( 2^k - 1 \) (split into single df), error \( (r - 1)(2^k - 1) \).
Yates’ Algorithm
For totals in standard order, form \( k \) new columns. In each, the first half is the sums of successive pairs of the previous column, and the second half is the differences (second minus first) of the same pairs. After \( k \) columns, the entries are \( G \) and then the contrasts \( [A], [B], [AB], [C], \dots \) in standard order.
Example (\( 2^3 \), \( r = 2 \), totals over the two replicates):
| Treatment | Total | Col 1 | Col 2 | Col 3 | Effect | Estimate \( \frac{[X]}{8} \) | SS \( \frac{[X]^2}{16} \) |
|---|---|---|---|---|---|---|---|
| (1) | 40 | 90 | 196 | 402 | G | ||
| a | 50 | 106 | 206 | 54 | A | 6.75 | 182.25 |
| b | 44 | 94 | 28 | 34 | B | 4.25 | 72.25 |
| ab | 62 | 112 | 26 | 18 | AB | 2.25 | 20.25 |
| c | 43 | 10 | 16 | 10 | C | 1.25 | 6.25 |
| ac | 51 | 18 | 18 | −2 | AC | −0.25 | 0.25 |
| bc | 47 | 8 | 8 | 2 | BC | 0.25 | 0.25 |
| abc | 65 | 18 | 10 | 2 | ABC | 0.25 | 0.25 |
Check: the seven SS add to 281.75, which equals \( \frac{1}{2}\sum(\text{totals})^2 - \frac{402^2}{16} \). Another check: the sum of squares of the last column equals \( 2^k \) times the sum of squares of the first.
Confounding
When the block size \( 2^{k-p} \) is smaller than \( 2^k \), some effects must be made identical to block contrasts. An effect is confounded with blocks when its contrast cannot be separated from block differences.
Two blocks: one confounded effect
To confound \( X \), put in one block all combinations with an even number of letters in common with \( X \), and in the other those with an odd number. Equivalently, with \( x_i \in \{0, 1\} \) the level of factor \( i \), split by \( L_X = \sum_{i \in X}x_i \pmod 2 \).
The block containing \( (1) \) is the principal block. For \( 2^3 \) with \( ABC \) confounded: principal block \( \{(1), ab, ac, bc\} \), other block \( \{a, b, c, abc\} \) (this is the hook’s split).
More blocks: generalized interactions
For \( 2^p \) blocks of size \( 2^{k-p} \), choose \( p \) independent effects to confound. Then every product of them (multiply letters, deleting squares: \( A^2 = I \)) is also confounded. The generalized interaction of \( X \) and \( Y \) is \( XY \) with squared letters removed. In all,
$$2^p - 1 \text{ effects are confounded with blocks (the block df).}$$Example. \( 2^4 \) in 4 blocks of 4, confounding \( ABC \) and \( BCD \): their product \( AB^2C^2D = AD \) is confounded too.
| Block (\( L_{ABC}, L_{BCD} \)) | Treatments |
|---|---|
| (0, 0), principal | (1), bc, abd, acd |
| (0, 1) | d, bcd, ab, ac |
| (1, 0) | a, abc, bd, cd |
| (1, 1) | b, c, ad, abcd |
Principal block is a subgroup. Under the product “multiply and delete squares” (addition in \( \mathbb Z_2^k \)), the principal block is closed: \( bc \cdot abd = ab^2cd = acd \). Every other block is a coset: multiply the principal block by any member of that block. So you only need the principal block plus one element of each other block.
graph TD
A["Block size 2^(k-p) needed"] --> B["Choose p independent effects, high-order first"]
B --> C["Add all generalized interactions: 2^p - 1 confounded"]
C --> D{Any main effect or key 2-factor interaction among them?}
D -->|Yes| B
D -->|No| E["Principal block: even overlap with every chosen effect"]
E --> F["Other blocks: cosets, multiply by one new treatment"]Choosing what to confound. Keep main effects and two-factor interactions clear when possible. In a \( 2^4 \) in 4 blocks, it is impossible to avoid confounding some two-factor interaction: if both chosen effects have three or more letters (out of four), their generalized interaction has at most two letters (e.g. \( ABC \cdot ABD = CD \), \( ABC \cdot ABCD = D \)). The best choice confounds two three-factor interactions and one two-factor interaction. In a \( 2^5 \) in 4 blocks of 8, \( ABC \) and \( CDE \) give \( ABDE \): only high-order interactions are lost.
Total confounding: the ANOVA
If the same effects are confounded in every replicate, they cannot be estimated at all. For \( 2^3 \) in blocks of 4 with \( ABC \) confounded in all \( r \) replicates:
| Source | df |
|---|---|
| Blocks | \( 2r - 1 \) (includes replicates and \( ABC \)) |
| A, B, C, AB, AC, BC | 1 each, 6 in all |
| Error | \( 6(r - 1) \) |
| Total | \( 8r - 1 \) |
Partial Confounding
Confounding different effects in different replicates keeps some information on every effect. Example: \( 2^3 \), 4 replicates, blocks of 4, confounding \( ABC \) in replicate I, \( AB \) in II, \( AC \) in III and \( BC \) in IV.
- Main effects \( A, B, C \) are never confounded: full information, computed from all 4 replicates.
- Each interaction is estimated only from the 3 replicates where it is not confounded. Relative information \( = \frac{3}{4} \).
- Its SS is \( \frac{[X]'^2}{r'2^k} \) with \( [X]' \) the contrast from the \( r' = 3 \) unconfounded replicates: \( \frac{[X]'^2}{24} \).
- Variance of the estimate: \( \frac{\sigma^2}{r'2^{k-2}} = \frac{\sigma^2}{6} \), against \( \frac{\sigma^2}{8} \) without confounding.
| Source | df |
|---|---|
| Blocks | 7 (8 blocks) |
| A, B, C, AB, AC, BC, ABC | 7 (each 1) |
| Error | 17 |
| Total | 31 |
In general, relative information on an effect is \( \frac{\text{number of replicates in which it is unconfounded}}{r} \).
Memory Tricks
Don’t Confuse
| This | vs | That |
|---|---|---|
| Treatment combination \( ab \) (a cell, or its total) | effect \( AB \) (a contrast among all cells) | |
| Total confounding: effect lost in every replicate | partial confounding: effect estimated from the replicates where it is not confounded | |
| Confounded effects \( = 2^p - 1 \) (block df within a replicate) | number of chosen “independent” effects \( = p \) | |
| Effect divisor \( r2^{k-1} \) | SS divisor \( r2^k \) | |
| Principal block (subgroup, contains (1)) | other blocks (cosets) | |
| Confounding (with blocks, all \( 2^k \) runs made) | fractional replication (only a fraction of runs; aliasing) |
Common Mistakes
Practice
All questions below are original practice questions written for this page. They are not from any CSIR NET paper. In Level 3, as in Part C, one or more options may be correct.
Level 1: Foundation
- How many treatment combinations and how many single-df factorial effects does a \( 2^4 \) experiment have?
- In a \( 2^3 \) experiment in 4 blocks of 2, \( AB \) and \( BC \) are confounded. Which other effect is confounded?
- Write the principal block of a \( 2^3 \) experiment in 2 blocks with \( ABC \) confounded.
Answers
- 16 combinations; 15 effects (4 main, 6 two-factor, 4 three-factor, 1 four-factor).
- \( AB \cdot BC = AB^2C = AC \).
- \( \{(1), ab, ac, bc\} \): each shares an even number (0 or 2) of letters with \( ABC \).
Level 2: Part B Style
Practice Q1. In a \( 2^2 \) experiment with 3 replicates, the treatment totals are \( (1) = 30 \), \( a = 42 \), \( b = 36 \), \( ab = 60 \). The main effect of \( A \) is
(a) 3 (b) 6 (c) 12 (d) 36
Answer: (b)
\( [A] = ab + a - b - (1) = 60 + 42 - 36 - 30 = 36 \). \( A = \frac{[A]}{r2^{k-1}} = \frac{36}{6} = 6 \). Option (a) uses the divisor \( r2^k = 12 \); (d) is the contrast itself. Also \( SS_A = \frac{36^2}{12} = 108 \), and \( [AB] = 60 - 42 - 36 + 30 = 12 \), so \( AB = 2 \).
Practice Q2. In the Yates table of this page (\( 2^3 \), \( r = 2 \)), the contrast for \( AB \) is 18. The sum of squares due to \( AB \) is
(a) 2.25 (b) 20.25 (c) 40.5 (d) 81
Answer: (b)
\( SS_{AB} = \frac{[AB]^2}{r2^k} = \frac{324}{16} = 20.25 \). Option (c) uses divisor 8 (\( r2^{k-1} \)); (a) is the effect estimate \( \frac{18}{8} \).
Practice Q3. A \( 2^4 \) experiment is arranged in 4 blocks of 4 with \( ABC \) and \( BCD \) confounded. The third confounded effect is
(a) \( AD \) (b) \( ABCD \) (c) \( BC \) (d) \( A \)
Answer: (a)
\( ABC \cdot BCD = AB^2C^2D = AD \). Option (c) takes the common letters instead of the symmetric difference.
Practice Q4. A \( 2^3 \) factorial is run in 4 replicates, each in 2 blocks of 4, with \( ABC, AB, AC, BC \) confounded in replicates I to IV respectively. The error df is
(a) 14 (b) 17 (c) 21 (d) 24
Answer: (b)
Total \( 32 - 1 = 31 \); blocks \( 8 - 1 = 7 \); treatments 7 (all seven effects remain estimable); error \( 31 - 7 - 7 = 17 \). Option (c) is the RBD value \( (4 - 1)(8 - 1) \).
Level 3: Part C Style
Practice Q5. A \( 2^5 \) experiment in factors \( A, \dots, E \) is arranged in 4 blocks of 8 by confounding \( ABC \) and \( CDE \). Which statements are true?
(a) \( ABDE \) is also confounded with blocks. (b) The principal block contains \( ab \), \( de \) and \( acd \). (c) No main effect or two-factor interaction is confounded. (d) The treatments \( a \) and \( b \) lie in the same block.
Answer: (a), (b), (c) and (d)
- (a) True. \( ABC \cdot CDE = ABC^2DE = ABDE \).
- (b) True. \( ab \): 2 letters in common with \( ABC \), 0 with \( CDE \). \( de \): 0 and 2. \( acd \): 2 (\( a, c \)) and 2 (\( c, d \)). All even. The full principal block is \( (1), ab, de, ace, acd, bce, bcd, abde \).
- (c) True. The confounded effects \( ABC, CDE, ABDE \) all have at least three letters.
- (d) True. \( a \): \( (L_{ABC}, L_{CDE}) = (1, 0) \); \( b \): \( (1, 0) \). Same block (indeed \( a \cdot b = ab \) is in the principal block, so they lie in the same coset).
Practice Q6. For the Yates example on this page (\( 2^3 \), \( r = 2 \), totals \( 40, 50, 44, 62, 43, 51, 47, 65 \) in standard order), which statements are true?
(a) The main effect of \( A \) is 6.75. (b) The treatment SS (7 df) is 281.75. (c) If the experiment had been run with \( ABC \) totally confounded, the treatment SS available for testing would be 281.5 on 6 df. (d) The estimated main effect of \( C \) has variance \( \frac{\sigma^2}{2} \).
Answer: (a), (b) and (c)
- (a) True. \( \frac{54}{2 \times 4} = 6.75 \).
- (b) True. \( 182.25 + 72.25 + 20.25 + 6.25 + 0.25 + 0.25 + 0.25 = 281.75 \).
- (c) True. Removing \( SS_{ABC} = 0.25 \) (now part of blocks) leaves 281.5 on 6 df.
- (d) False. \( V = \frac{\sigma^2}{r2^{k-2}} = \frac{\sigma^2}{2 \times 2} = \frac{\sigma^2}{4} \).
Practice Q7. In the partially confounded \( 2^3 \) of Q4 (\( ABC, AB, AC, BC \) confounded in replicates I to IV), which statements are true?
(a) The relative information on \( AB \) is \( \frac{3}{4} \). (b) The relative information on \( A \) is 1. (c) The variance of the estimate of \( AC \) is \( \frac{\sigma^2}{6} \). (d) \( SS_{BC} = \frac{[BC]^2}{32} \), where \( [BC] \) is computed from all four replicates.
Answer: (a), (b) and (c)
- (a) True. \( AB \) is unconfounded in 3 of the 4 replicates.
- (b) True. Main effects are never confounded here.
- (c) True. \( \frac{\sigma^2}{r'2^{k-2}} = \frac{\sigma^2}{3 \times 2} = \frac{\sigma^2}{6} \).
- (d) False. \( BC \) is confounded in replicate IV, so \( [BC]' \) uses replicates I to III and \( SS_{BC} = \frac{[BC]'^2}{3 \times 8} = \frac{[BC]'^2}{24} \).
Quick Revision Table
| Idea | Fact |
|---|---|
| Contrast | \( [X] = \prod(\ell \pm 1) \), minus for letters in \( X \), \( (1) \) for the constant |
| Effect | \( \frac{[X]}{r2^{k-1}} \); variance \( \frac{\sigma^2}{r2^{k-2}} \) |
| SS | \( \frac{[X]^2}{r2^k} \), 1 df each; \( 2^k - 1 \) of them sum to treatment SS |
| \( 2^2 \) | \( A = \frac{ab + a - b - (1)}{2r} \), \( AB = \frac{ab - a - b + (1)}{2r} \) |
| Yates | \( k \) columns: pair sums then pair differences |
| Confounding into \( 2^p \) blocks | choose \( p \) effects; all \( 2^p - 1 \) products confounded |
| Generalized interaction | multiply, delete squares: \( ABC \cdot BCD = AD \) |
| Principal block | even overlap with every confounded effect; a subgroup |
| \( 2^4 \) in 4 blocks | some 2-factor interaction must be confounded |
| Partial confounding | relative information \( \frac{r'}{r} \); SS from unconfounded replicates, divisor \( r'2^k \) |
Key Takeaways
What’s Next
- Sampling and Design of Experiments: Quick Revision: every formula of the chapter on one page.
- Block Designs: Connectedness, Orthogonality and BIBD: the other answer to “blocks too small”.
- Fields, Extensions and Galois Theory: the arithmetic of \( GF(2) \) behind “delete squares” and of \( GF(p^k) \) behind orthogonal Latin squares.
- Back to the Mathematical Sciences overview.