Block Designs: Connectedness, Orthogonality and BIBD
CSIR NET block designs: incidence and C-matrix, connectedness, orthogonality, BIBD parameter relations, Fisher's inequality and intrablock analysis.
- 20 min
On this page · 13 sections
Prerequisites
- CRD, RBD and Latin Square Designs: the RBD is the complete, orthogonal block design that everything here generalises.
- Rank, Determinants and Linear Equations: estimability is a statement about the row space and rank of the C-matrix.
- Eigenvalues and Cayley-Hamilton: the eigenvalues of \( aI + bJ \) give \( \det NN' \) and Fisher’s inequality in one line.
- Combinatorics: Counting, Pigeonhole and Derangements: the BIBD relations are double-counting arguments.
The Hook: Seven Wines, Three Glasses
The General Block Design
\( v \) treatments are applied in \( b \) blocks. The incidence matrix \( N = (n_{ij}) \) is \( v \times b \), with \( n_{ij} \) the number of times treatment \( i \) occurs in block \( j \).
| Symbol | Meaning |
|---|---|
| \( r_i = \sum_j n_{ij} \), \( R = \operatorname{diag}(r_i) \) | replication of treatment \( i \) |
| \( k_j = \sum_i n_{ij} \), \( K = \operatorname{diag}(k_j) \) | size of block \( j \) |
| \( n = \sum r_i = \sum k_j \) | number of plots |
| \( T_i \), \( B_j \), \( G \) | treatment totals, block totals, grand total |
The design is binary if every \( n_{ij} \in \{0, 1\} \), proper if all \( k_j = k \), equireplicate if all \( r_i = r \).
The reduced normal equations
Under the model \( y_{ijl} = \mu + \tau_i + \beta_j + e_{ijl} \), eliminating \( \mu \) and the \( \beta_j \) from the least-squares equations leaves
$$\boxed{C\hat\tau = Q, \qquad C = R - NK^{-1}N', \qquad Q = T - NK^{-1}B.}$$\( Q_i = T_i - \sum_j \frac{n_{ij}B_j}{k_j} \) is the adjusted treatment total: treatment \( i \)’s total minus its share of the totals of the blocks it sits in. Properties of the C-matrix (information matrix):
- \( C \) is symmetric and positive semidefinite, with every row sum zero (\( C\mathbf 1 = 0 \)), so \( \operatorname{rank} C \le v - 1 \).
- \( E(Q) = C\tau \) and \( D(Q) = \sigma^2C \); \( \sum Q_i = 0 \).
- A linear function \( \ell'\tau \) is estimable iff \( \ell \) lies in the row space of \( C \). Then its BLUE is \( \ell'\hat\tau \) for any solution \( \hat\tau \), with variance \( \sigma^2\ell'C^-\ell \) for any generalised inverse \( C^- \).
- The adjusted treatment SS is \( \hat\tau'Q \) on \( \operatorname{rank} C \) df; error df \( = n - b - \operatorname{rank} C \).
Connectedness
Two treatments are connected if one can pass from one to the other by a chain treatment, block, treatment, block, …, where consecutive members occur together. The design is connected if every pair of treatments is connected.
$$\boxed{\text{connected} \iff \operatorname{rank} C = v - 1 \iff \text{every treatment contrast is estimable}.}$$Why. Off the diagonal, \( c_{ii'} = -\sum_j \frac{n_{ij}n_{i'j}}{k_j} \le 0 \), and rows sum to 0. So \( C \) is the Laplacian of a weighted graph on the treatments, with an edge between two treatments whenever they share a block. A graph Laplacian has rank \( v - (\text{number of connected components}) \); it is \( v - 1 \) exactly when the graph is connected. The null vector \( \mathbf 1 \) is then the only direction lost, and \( \ell'\tau \) is estimable iff \( \ell'\mathbf 1 = 0 \), i.e. iff it is a contrast.
Necessary condition (binary designs). The treatment-block incidence graph has \( v + b \) vertices and \( n \) edges; to be connected it needs \( n \ge v + b - 1 \).
Three small designs on \( v = 4 \)
| Blocks | \( C \) (rows) | rank | Verdict |
|---|---|---|---|
| \( \{1, 2\}, \{2, 3\}, \{3, 4\} \) | \( (\tfrac12, -\tfrac12, 0, 0), (-\tfrac12, 1, -\tfrac12, 0), (0, -\tfrac12, 1, -\tfrac12), (0, 0, -\tfrac12, \tfrac12) \) | 3 | connected (a chain) |
| \( \{1, 2\}, \{1, 2\}, \{3, 4\}, \{3, 4\} \) | \( (1, -1, 0, 0), (-1, 1, 0, 0), (0, 0, 1, -1), (0, 0, -1, 1) \) | 2 | disconnected: \( \tau_1 - \tau_3 \) not estimable |
| \( \{1, 2, 3\}, \{1, 2, 4\} \) | \( (\tfrac43, -\tfrac23, -\tfrac13, -\tfrac13), (-\tfrac23, \tfrac43, -\tfrac13, -\tfrac13), (-\tfrac13, -\tfrac13, \tfrac23, 0), (-\tfrac13, -\tfrac13, 0, \tfrac23) \) | 3 | connected, but unequal precision |
In the third design, \( V(\hat\tau_1 - \hat\tau_2) = \sigma^2 \), \( V(\hat\tau_1 - \hat\tau_3) = 1.75\sigma^2 \) and \( V(\hat\tau_3 - \hat\tau_4) = 3\sigma^2 \): treatments 3 and 4 never meet and are compared only through 1 and 2.
Orthogonality
A connected block design is orthogonal when treatment and block effects can be estimated independently: the adjusted and unadjusted treatment SS coincide, and block totals carry no information on treatment contrasts. The criterion is
$$\boxed{n_{ij} = \frac{r_ik_j}{n} \ \text{for all } i, j, \qquad \text{i.e.} \qquad N = \frac{1}{n}\,r\,k'.}$$Then \( C = R - \frac{1}{n}rr' \), \( Q_i = T_i - \frac{r_iG}{n} \), and \( \hat\tau_i - \hat\tau_{i'} = \bar y_i - \bar y_{i'} \): exactly as if there were no blocks.
Proof idea. \( Q = T - NK^{-1}B \). With \( N = \frac{rk'}{n} \), \( NK^{-1}B = \frac{r}{n}\mathbf 1'B = \frac{G}{n}r \), so \( Q \) is free of individual block totals. Conversely, if the design is orthogonal the block totals must enter \( Q \) only through \( G \), which forces the proportional pattern.
- RBD: \( n_{ij} = 1 = \frac{r \cdot t}{rt} \). Orthogonal.
- Non-binary orthogonal example: \( v = 3 \), \( b = 2 \), \( N = \begin{pmatrix} 2 & 1 \\ 2 & 1 \\ 2 & 1 \end{pmatrix} \): \( r_i = 3 \), \( k = (6, 3) \), \( n = 9 \), and \( \frac{3 \cdot 6}{9} = 2 \), \( \frac{3 \cdot 3}{9} = 1 \). Orthogonal.
- Every incomplete binary design is non-orthogonal: if some \( n_{ij} = 0 \), proportionality forces \( r_i = 0 \) or \( k_j = 0 \).
Balance and efficiency
A connected design is variance balanced if every elementary contrast \( \tau_i - \tau_{i'} \) is estimated with the same variance; for an equireplicate proper design this means \( C = \theta\left(I - \frac{1}{v}J\right) \). The efficiency factor compares that variance with \( \frac{2\sigma^2}{r} \), the RBD value at the same replication:
$$E = \frac{2\sigma^2/r}{V(\hat\tau_i - \hat\tau_{i'})}.$$Balanced Incomplete Block Designs (BIBD)
A BIBD with parameters \( (v, b, r, k, \lambda) \) is a binary design in which \( v \) treatments are arranged in \( b \) blocks of size \( k < v \), each treatment occurs in \( r \) blocks, and every pair of distinct treatments occurs together in exactly \( \lambda \) blocks.
Parameter relations
$$\boxed{bk = vr, \qquad \lambda(v - 1) = r(k - 1), \qquad b \ge v \ (\text{Fisher's inequality}).}$$- \( bk = vr \): count the \( n \) plots by blocks and by treatments.
- \( \lambda(v - 1) = r(k - 1) \): fix treatment \( \theta \). Count pairs (\( \phi \), block) with \( \phi \ne \theta \) and both in the block: \( \theta \) is in \( r \) blocks, each with \( k - 1 \) partners; each of the \( v - 1 \) partners shares \( \lambda \) blocks with \( \theta \).
- \( NN' = (r - \lambda)I + \lambda J \): diagonal \( r \), off-diagonal \( \lambda \).
- Fisher’s inequality. \( NN' \) has eigenvalues \( r - \lambda + \lambda v = rk \) (once, on \( \mathbf 1 \)) and \( r - \lambda \) (\( v - 1 \) times), so \( \det NN' = rk(r - \lambda)^{v - 1} \). Since \( k < v \) gives \( r > \lambda \), \( NN' \) is non-singular, \( v = \operatorname{rank}NN' \le \operatorname{rank}N \le b \). Then also \( r \ge k \).
Intrablock analysis
From \( NN' \), with \( r(k - 1) + \lambda = \lambda v \):
$$C = rI - \frac{1}{k}NN' = \frac{\lambda v}{k}\left(I - \frac{1}{v}J\right).$$So \( \operatorname{rank} C = v - 1 \) (connected), and with \( \sum\hat\tau_i = 0 \):
$$\boxed{\hat\tau_i = \frac{kQ_i}{\lambda v}, \qquad V(\hat\tau_i - \hat\tau_{i'}) = \frac{2k\sigma^2}{\lambda v}, \qquad E = \frac{\lambda v}{rk} = \frac{1 - 1/k}{1 - 1/v} < 1.}$$| Source | df | SS |
|---|---|---|
| Blocks (unadjusted) | \( b - 1 \) | \( \frac{1}{k}\sum B_j^2 - CF \) |
| Treatments (adjusted) | \( v - 1 \) | \( \frac{k}{\lambda v}\sum Q_i^2 \) |
| Error | \( bk - b - v + 1 \) | by subtraction |
| Total | \( bk - 1 \) | \( \sum y^2 - CF \) |
The design of the hook, \( (7, 7, 3, 3, 1) \): \( V(\hat\tau_i - \hat\tau_j) = \frac{2 \cdot 3}{1 \cdot 7}\sigma^2 = \frac{6}{7}\sigma^2 \), \( E = \frac{7}{9} \).
Worked example: all pairs of four treatments
The \( (4, 6, 3, 2, 1) \) design has blocks \( \{1, 2\}, \{1, 3\}, \{1, 4\}, \{2, 3\}, \{2, 4\}, \{3, 4\} \). Observations (first value for the lower-numbered treatment):
| Block | \( \{1,2\} \) | \( \{1,3\} \) | \( \{1,4\} \) | \( \{2,3\} \) | \( \{2,4\} \) | \( \{3,4\} \) |
|---|---|---|---|---|---|---|
| Yields | 10, 12 | 11, 15 | 9, 11 | 13, 16 | 12, 12 | 17, 14 |
| \( B_j \) | 22 | 26 | 20 | 29 | 24 | 31 |
Treatment totals \( T = (30, 37, 48, 37) \). Treatment 1 sits in blocks with totals 22, 26, 20, so \( Q_1 = 30 - \frac{68}{2} = -4 \); similarly \( Q = (-4, -0.5, 5, -0.5) \) (sum 0). Then \( \hat\tau = \frac{2Q}{1 \cdot 4} = \frac{Q}{2} = (-2, -0.25, 2.5, -0.25) \).
\( G = 152 \), \( CF = 1925.33 \): total SS \( = 64.67 \), blocks (unadjusted) \( = 43.67 \), treatments (adjusted) \( = \frac{2}{4}(16 + 0.25 + 25 + 0.25) = 20.75 \), error \( = 0.25 \) on \( 12 - 6 - 4 + 1 = 3 \) df. Note the unadjusted treatment SS is \( \frac{1}{3}\sum T_i^2 - CF = 55.33 \): very different, because the design is not orthogonal.
Families and existence
| Design | Parameters |
|---|---|
| Unreduced (all \( k \)-subsets) | \( b = \binom{v}{k} \), \( r = \binom{v-1}{k-1} \), \( \lambda = \binom{v-2}{k-2} \) |
| Complement of \( (v, b, r, k, \lambda) \) | \( (v, b, b - r, v - k, b - 2r + \lambda) \), valid when \( v - k \ge 2 \) |
| Symmetric | \( b = v \), \( r = k \); any two blocks share exactly \( \lambda \) treatments; \( \lambda(v - 1) = k(k - 1) \) |
| Resolvable | blocks split into \( r \) replicates each containing every treatment once; then \( b \ge v + r - 1 \) |
| Youden square | a symmetric BIBD laid out as a \( k \times v \) array whose columns are the blocks and in which every treatment occurs once in each row (a Latin square with \( v - k \) rows deleted) |
Existence. The relations are necessary, not sufficient. For a symmetric BIBD with \( v \) even, \( r - \lambda \) must be a perfect square (part of the Bruck-Ryser-Chowla theorem). Example: \( v = 22 \), \( k = 7 \), \( \lambda = 2 \) gives \( r = \frac{2 \cdot 21}{6} = 7 \), \( b = 22 \), symmetric with \( v \) even, but \( r - \lambda = 5 \) is not a square, so no such design exists.
graph TD
A[Candidate v, k, lambda] --> B["r = lambda(v-1)/(k-1) integer?"]
B -->|No| X[No BIBD]
B -->|Yes| C["b = vr/k integer?"]
C -->|No| X
C -->|Yes| D{"b >= v (Fisher)?"}
D -->|No| X
D -->|Yes| E{"Symmetric (b = v) with v even?"}
E -->|Yes| F{"r - lambda a perfect square?"}
F -->|No| X
F -->|Yes| G[Necessary conditions pass]
E -->|No| GMemory Tricks
Don’t Confuse
| This | vs | That |
|---|---|---|
| Connected: all contrasts estimable, rank \( C = v - 1 \) | orthogonal: treatment estimates unaffected by blocks, \( N = \frac{rk'}{n} \) | |
| Balanced (equal variance of all elementary contrasts) | orthogonal; a BIBD is balanced but not orthogonal | |
| \( Q_i \) adjusted total (block effects removed) | \( T_i \) raw total | |
| Adjusted treatment SS \( \frac{k}{\lambda v}\sum Q_i^2 \) (test treatments) | unadjusted treatment SS \( \frac{1}{r}\sum T_i^2 - CF \) (not valid for the test) | |
| Efficiency factor \( E = \frac{\lambda v}{rk} \) | relative efficiency in practice also depends on the smaller \( \sigma^2 \) in small blocks | |
| Symmetric BIBD (\( b = v \)) | resolvable BIBD (\( b \ge v + r - 1 \)) |
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
- Check that \( (v, b, r, k, \lambda) = (6, 10, 5, 3, 2) \) satisfies the BIBD relations.
- What is the rank of the C-matrix of a connected block design with \( v \) treatments?
- Find the efficiency factor of the \( (7, 7, 3, 3, 1) \) design.
Answers
- \( bk = 30 = vr \); \( \lambda(v - 1) = 10 = r(k - 1) \); \( b = 10 \ge v = 6 \).
- \( v - 1 \).
- \( E = \frac{\lambda v}{rk} = \frac{7}{9} \approx 0.778 \).
Level 2: Part B Style
Practice Q1. A BIBD has \( v = 9 \), \( k = 3 \), \( \lambda = 1 \). The number of blocks is
(a) 9 (b) 12 (c) 18 (d) 36
Answer: (b)
\( r = \frac{\lambda(v - 1)}{k - 1} = \frac{8}{2} = 4 \); \( b = \frac{vr}{k} = \frac{36}{3} = 12 \). Option (d) is \( vr \) without dividing by \( k \). This design (an affine plane of order 3) is resolvable, with \( b = 12 = v + r - 1 \).
Practice Q2. In the \( (7, 7, 3, 3, 1) \) BIBD with error variance \( \sigma^2 \), the variance of the intrablock estimate of \( \tau_1 - \tau_2 \) is
(a) \( \frac{2}{3}\sigma^2 \) (b) \( \frac{6}{7}\sigma^2 \) (c) \( \frac{7}{9}\sigma^2 \) (d) \( \frac{3}{7}\sigma^2 \)
Answer: (b)
\( \frac{2k\sigma^2}{\lambda v} = \frac{6}{7}\sigma^2 \). Option (a) is the RBD value \( \frac{2\sigma^2}{r} \); (c) is the efficiency factor; (d) is \( \frac{k\sigma^2}{\lambda v} \), half the correct value (the factor 2 of a difference is forgotten).
Practice Q3. Which of the following parameter sets cannot be those of a BIBD?
(a) \( v = 7, k = 3, \lambda = 1 \) (b) \( v = 6, k = 3, \lambda = 2 \) (c) \( v = 16, k = 6, \lambda = 1 \) (d) \( v = 9, k = 3, \lambda = 1 \)
Answer: (c)
(c): \( r = \frac{15}{5} = 3 \), \( b = \frac{16 \times 3}{6} = 8 < v = 16 \), violating Fisher’s inequality. The others give \( (7, 7, 3, 3, 1) \), \( (6, 10, 5, 3, 2) \) and \( (9, 12, 4, 3, 1) \), all of which exist.
Practice Q4. For a block design with \( v = 4 \) and blocks \( \{1, 2\}, \{1, 2\}, \{3, 4\}, \{3, 4\} \), which contrast is not estimable?
(a) \( \tau_1 - \tau_2 \) (b) \( \tau_3 - \tau_4 \) (c) \( \tau_1 + \tau_2 - \tau_3 - \tau_4 \) (d) \( 2\tau_1 - 2\tau_2 + \tau_3 - \tau_4 \)
Answer: (c)
The design splits into two disconnected parts \( \{1, 2\} \) and \( \{3, 4\} \); \( \operatorname{rank}C = 2 \). The row space of \( C \) is spanned by \( (1, -1, 0, 0) \) and \( (0, 0, 1, -1) \). (a), (b) and (d) are combinations of these; (c) compares the two parts, which never meet, and is confounded with block effects.
Level 3: Part C Style
Practice Q5. A design has \( v = 4 \) treatments in two blocks \( \{1, 2, 3\} \) and \( \{1, 2, 4\} \). Which statements are true?
(a) The design is connected. (b) The design is orthogonal. (c) \( V(\hat\tau_3 - \hat\tau_4) = 3\sigma^2 \). (d) \( V(\hat\tau_1 - \hat\tau_2) = \sigma^2 \).
Answer: (a), (c) and (d)
- (a) True. 3 meets 1 and 2 in block 1, 4 meets them in block 2; \( \operatorname{rank}C = 3 = v - 1 \). (Also \( n = 6 \ge v + b - 1 = 5 \).)
- (b) False. \( r = (2, 2, 1, 1) \), \( k = (3, 3) \), \( n = 6 \): \( \frac{r_3k_1}{n} = \frac{1}{2} \ne n_{31} = 1 \).
- (c) True. The best estimate is \( (y_{3} - \frac{y_{11} + y_{21}}{2}) - (y_{4} - \frac{y_{12} + y_{22}}{2}) \) (within-block deviations), with variance \( 2 \times \sigma^2(1 + \frac{1}{2}) = 3\sigma^2 \); equivalently \( \ell'C^-\ell = 3 \) for \( \ell = (0, 0, 1, -1) \).
- (d) True. Treatments 1 and 2 meet in both blocks; the two within-block differences each have variance \( 2\sigma^2 \), and their average has variance \( \sigma^2 \).
Practice Q6. Which statements about BIBDs are true?
(a) A symmetric BIBD with parameters \( (22, 22, 7, 7, 2) \) exists. (b) The complement of the \( (7, 7, 3, 3, 1) \) design is a \( (7, 7, 4, 4, 2) \) BIBD. (c) In every BIBD, \( r \ge k \). (d) In every BIBD, the efficiency factor \( \frac{\lambda v}{rk} \) is strictly less than 1.
Answer: (b), (c) and (d)
- (a) False. \( v = 22 \) is even and \( r - \lambda = 5 \) is not a perfect square (Bruck-Ryser-Chowla).
- (b) True. \( b - r = 4 \), \( v - k = 4 \), \( b - 2r + \lambda = 7 - 6 + 1 = 2 \).
- (c) True. From \( bk = vr \) and \( b \ge v \): \( r = \frac{bk}{v} \ge k \).
- (d) True. \( E = \frac{1 - 1/k}{1 - 1/v} \) and \( k < v \).
Practice Q7. For the \( (4, 6, 3, 2, 1) \) data in the worked example (\( T = (30, 37, 48, 37) \), block totals 22, 26, 20, 29, 24, 31), which statements are true?
(a) \( Q_3 = 5 \). (b) The intrablock estimate of \( \tau_3 - \tau_1 \) is 4.5. (c) The adjusted treatment SS equals the unadjusted treatment SS, 55.33. (d) The error has 3 df.
Answer: (a), (b) and (d)
- (a) True. Treatment 3 is in blocks \( \{1,3\}, \{2,3\}, \{3,4\} \) with totals 26, 29, 31: \( Q_3 = 48 - \frac{86}{2} = 5 \).
- (b) True. \( \frac{k(Q_3 - Q_1)}{\lambda v} = \frac{2(5 - (-4))}{4} = 4.5 \). (The raw means differ by \( 16 - 10 = 6 \).)
- (c) False. Adjusted SS \( = \frac{2}{4}\sum Q_i^2 = 20.75 \); the unadjusted 55.33 includes block differences, because the design is not orthogonal.
- (d) True. \( bk - b - v + 1 = 12 - 6 - 4 + 1 = 3 \).
Quick Revision Table
| Idea | Fact |
|---|---|
| C-matrix | \( C = R - NK^{-1}N' \), \( Q = T - NK^{-1}B \), \( C\hat\tau = Q \) |
| Properties | symmetric, psd, row sums 0, rank \( \le v - 1 \); \( D(Q) = \sigma^2C \) |
| Estimable | \( \ell'\tau \) iff \( \ell \in \) row space of \( C \); variance \( \sigma^2\ell'C^-\ell \) |
| Connected | rank \( C = v - 1 \) iff all contrasts estimable; binary needs \( n \ge v + b - 1 \) |
| Orthogonal | \( n_{ij} = \frac{r_ik_j}{n} \); then \( \hat\tau_i - \hat\tau_{i'} = \bar y_i - \bar y_{i'} \) |
| BIBD relations | \( bk = vr \), \( \lambda(v-1) = r(k-1) \), \( b \ge v \), \( r \ge k \), \( r > \lambda \) |
| \( NN' \) | \( (r - \lambda)I + \lambda J \); det \( rk(r - \lambda)^{v-1} \) |
| BIBD analysis | \( C = \frac{\lambda v}{k}(I - \frac{J}{v}) \); \( \hat\tau = \frac{kQ}{\lambda v} \); \( V = \frac{2k\sigma^2}{\lambda v} \); \( E = \frac{\lambda v}{rk} \) |
| Error df | \( bk - b - v + 1 \) |
| Complement | \( (v, b, b - r, v - k, b - 2r + \lambda) \) |
| Symmetric | \( b = v \), \( r = k \); blocks meet in \( \lambda \); \( v \) even \( \Rightarrow r - \lambda \) square |
Key Takeaways
What’s Next
- 2^k Factorial Experiments and Confounding: when blocks are too small for a full factorial, confounding deliberately sacrifices chosen interactions, a different way to live with incomplete blocks.
- CRD, RBD and Latin Square Designs: the complete-block designs, for comparison.
- Sampling and Design of Experiments: Quick Revision.