Chapter 8 · Sample Surveys and Design of Experiments

8.4 Block Designs, BIBD and 2ᵏ Factorial Experiments

Designs whose blocks are too small to hold every treatment: the incidence matrix and the C-matrix, connectedness, completeness and orthogonality; balanced incomplete block designs with their parameter relations, Fisher's inequality, symmetric designs and the intrablock analysis; and 2ᵏ factorial experiments with Yates' algorithm, total and partial confounding and the construction of blocks.

A wine panel can taste perhaps four wines before the palate tires, but there are seven wines to compare. A tyre manufacturer wants to compare five rubber compounds, but a car has only four wheels. An engineer studies four factors at two levels each — sixteen combinations — but the furnace takes only eight specimens a day. In each case the natural block (a taster's session, a car, a day) is smaller than the set of treatments, and the randomised block design of §8.3 is impossible. Two classical answers are developed in this section. If all treatments matter equally, use an incomplete block design in which every pair of treatments meets equally often — a balanced incomplete block design (BIBD). If the treatments are the combinations of several factors, and some high-order interactions are expected to be negligible, sacrifice those interactions on purpose and confound them with blocks.

Both rest on one piece of linear algebra. Eliminating the block effects from the normal equations of §6.1 leaves the reduced normal equations $C\hat\tau=Q$, and the properties of the C-matrix decide everything: which treatment contrasts are estimable (connectedness), whether blocks can be ignored (orthogonality), and how precisely contrasts are estimated (balance and efficiency). We then study BIBDs — their parameter relations $bk=vr$ and $\lambda(v-1)=r(k-1)$, Fisher's inequality $b\ge v$, symmetric designs and their intrablock analysis — and finally $2^k$ factorial experiments: main effects and interactions as orthogonal contrasts, Yates' algorithm, and total and partial confounding with the construction of blocks by arithmetic modulo $2$. CSIR NET Part B asks for a missing BIBD parameter, a generalised interaction or an efficiency factor; Part C asks which statements about connectedness, Fisher's inequality, symmetric designs and confounding are true. Notation follows §6.1 ($^{\mathsf T}$ for transpose, $\mathcal C(A)$ for a column space, $\mathbf 1$ and $J=\mathbf 1\mathbf 1^{\mathsf T}$).

Block Designs and the C-Matrix

📘 Definition 8.15 — Block design and incidence matrix

A block design allots $v$ treatments to the plots of $b$ blocks, block $j$ having $k_j$ plots. Its incidence matrix is the $v\times b$ matrix $N=(n_{ij})$, where $n_{ij}$ is the number of plots of block $j$ receiving treatment $i$. Then the replication of treatment $i$ is $r_i=\sum_jn_{ij}$, the block size is $k_j=\sum_in_{ij}$, and $n=\sum_ir_i=\sum_jk_j$ is the number of plots. Write $r=(r_1,\dots,r_v)^{\mathsf T}$, $k=(k_1,\dots,k_b)^{\mathsf T}$, $R=\operatorname{diag}(r)$, $K=\operatorname{diag}(k)$, so that $N\mathbf 1_b=r$ and $N^{\mathsf T}\mathbf 1_v=k$. The design is

  • binary if every $n_{ij}\in\{0,1\}$ (no treatment twice in a block); proper if all $k_j=k$; equireplicate if all $r_i=r$;
  • complete if every block contains every treatment ($n_{ij}\ge1$), and incomplete otherwise.

The model is $y_{ijm}=\mu+\tau_i+\beta_j+\varepsilon_{ijm}$ for the plots of block $j$ with treatment $i$; in matrix form $y=\mu\mathbf 1+X_\tau\tau+X_\beta\beta+\varepsilon$ with $X_\tau^{\mathsf T}X_\tau=R$, $X_\beta^{\mathsf T}X_\beta=K$ and $X_\tau^{\mathsf T}X_\beta=N$. Let $T=X_\tau^{\mathsf T}y$ and $B=X_\beta^{\mathsf T}y$ be the vectors of treatment and block totals.

The RBD of §8.3 is the binary, proper, equireplicate, complete design with $N=J_{v\times b}$. In an incomplete design a treatment total $T_i$ is contaminated by the effects of the blocks in which treatment $i$ happens to lie, and simple treatment means are biased. The remedy is to adjust the totals for blocks.

📐 Theorem 8.29 — Reduced normal equations and the C-matrix

Define the C-matrix and the adjusted treatment totals $$C=R-NK^{-1}N^{\mathsf T},\qquad Q=T-NK^{-1}B,$$ so that $Q_i=T_i-\sum_jn_{ij}B_j/k_j$ (the treatment total minus the average of the block totals in which it occurs, counted with multiplicity). Then:

  1. the least squares estimates of $\tau$ are the solutions of the reduced normal equations $C\hat\tau=Q$;
  2. $E[Q]=C\tau$ and $\operatorname{Cov}(Q)=\sigma^2C$; block effects have disappeared;
  3. $C$ is symmetric, positive semidefinite, with zero row sums ($C\mathbf 1=0$), so $\operatorname{rank}C\le v-1$;
  4. $\lambda^{\mathsf T}\tau$ is estimable iff $\lambda\in\mathcal C(C)$; its BLUE is $\lambda^{\mathsf T}\hat\tau$, with variance $\sigma^2\lambda^{\mathsf T}C^-\lambda$ for any generalised inverse $C^-$;
  5. the sum of squares for treatments adjusted for blocks is $\hat\tau^{\mathsf T}Q$ on $\operatorname{rank}C$ df, and the error sum of squares is $\text{SST}-\text{SSB}_{\text{unadj}}-\hat\tau^{\mathsf T}Q$ on $n-b-\operatorname{rank}C$ df, where $\text{SSB}_{\text{unadj}}=\sum_jB_j^2/k_j-G^2/n$.

Proof. Since $X_\beta\mathbf 1_b=\mathbf 1_n$, the column $\mathbf 1$ is in $\mathcal C(X_\beta)$ and $\mu$ can be absorbed into $\beta$. The normal equations (§6.1) for $(\tau,\beta)$ are $$R\hat\tau+N\hat\beta=T,\qquad N^{\mathsf T}\hat\tau+K\hat\beta=B.$$ The second gives $\hat\beta=K^{-1}(B-N^{\mathsf T}\hat\tau)$; substituting in the first gives $(R-NK^{-1}N^{\mathsf T})\hat\tau=T-NK^{-1}B$, which is (1). For (2) let $P_\beta=X_\beta K^{-1}X_\beta^{\mathsf T}$, the orthogonal projection onto the block space (it replaces each observation by its block mean). Then $Q=X_\tau^{\mathsf T}(I-P_\beta)y$, and $(I-P_\beta)$ annihilates $\mu\mathbf 1+X_\beta\beta$, so $E[Q]=X_\tau^{\mathsf T}(I-P_\beta)X_\tau\tau=C\tau$ and $\operatorname{Cov}(Q)=\sigma^2X_\tau^{\mathsf T}(I-P_\beta)X_\tau=\sigma^2C$, because $I-P_\beta$ is symmetric and idempotent. This also shows $C=Z^{\mathsf T}Z$ with $Z=(I-P_\beta)X_\tau$, hence (3): $C$ is symmetric and positive semidefinite, and $C\mathbf 1=r-NK^{-1}k=r-N\mathbf 1_b=0$.

(4) Multiplying $y$ by $I-P_\beta$ gives the model $(I-P_\beta)y=Z\tau+(I-P_\beta)\varepsilon$, which contains all the information about $\tau$ that is free of the blocks; by the estimability criterion of §6.1 the estimable functions are those with $\lambda$ in the row space of $Z$, which equals $\mathcal C(Z^{\mathsf T}Z)=\mathcal C(C)$. If $\lambda=Cu$ then $\lambda^{\mathsf T}\hat\tau=u^{\mathsf T}C\hat\tau=u^{\mathsf T}Q$ does not depend on which solution is used, is unbiased by (2), and has variance $\sigma^2u^{\mathsf T}Cu=\sigma^2\lambda^{\mathsf T}C^-\lambda$; it is the least squares estimate and hence the BLUE (Gauss–Markov, §6.1). (5) The reduction in the residual sum of squares obtained by adding treatments to the blocks-only model is $\|P_Zy\|^2=\hat\tau^{\mathsf T}Z^{\mathsf T}y=\hat\tau^{\mathsf T}Q$, with df $\operatorname{rank}Z=\operatorname{rank}C$; the blocks-only model has residual $\text{SST}-\text{SSB}_{\text{unadj}}$ on $n-b$ df. ∎

📘 Definition 8.16 — Connected design
A treatment $i$ and a block $j$ are associated if $n_{ij}\ge1$. Two treatments are connected if there is a chain treatment–block–treatment–block–…–treatment from one to the other in which consecutive members are associated. The design is connected if every pair of treatments is connected; equivalently, the bipartite graph with treatments and blocks as vertices and an edge for each $n_{ij}\ge1$ is connected.
📐 Theorem 8.30 — Connectedness and the rank of C

For every $\tau\in\mathbb R^v$, $$\tau^{\mathsf T}C\tau=\sum_{j=1}^b\frac1{k_j}\sum_{i\lt i'}n_{ij}n_{i'j}(\tau_i-\tau_{i'})^2.$$ Consequently $\operatorname{rank}C=v-c$, where $c$ is the number of connected components of the design, and the following are equivalent: (a) the design is connected; (b) $\operatorname{rank}C=v-1$; (c) every treatment contrast $\sum\lambda_i\tau_i$ ($\sum\lambda_i=0$) is estimable.

Proof. The contribution of block $j$ to $\tau^{\mathsf T}C\tau=\tau^{\mathsf T}R\tau-\tau^{\mathsf T}NK^{-1}N^{\mathsf T}\tau$ is $\sum_in_{ij}\tau_i^2-\frac1{k_j}\left(\sum_in_{ij}\tau_i\right)^2$, which is $k_j$ times the variance of the values $\tau_i$ weighted by $n_{ij}/k_j$. Expanding the square and using $\sum_in_{ij}=k_j$, $$\frac1{2k_j}\sum_{i,i'}n_{ij}n_{i'j}(\tau_i-\tau_{i'})^2=\frac{1}{2k_j}\Bigl[2k_j\sum_in_{ij}\tau_i^2-2\Bigl(\sum_in_{ij}\tau_i\Bigr)^2\Bigr],$$ which is the same quantity; the left side is the stated double sum over $i\lt i'$. Since $C$ is positive semidefinite, $C\tau=0$ iff $\tau^{\mathsf T}C\tau=0$, i.e. iff $\tau_i=\tau_{i'}$ whenever $i,i'$ share a block, i.e. iff $\tau$ is constant on each connected component. So the null space of $C$ is spanned by the $c$ indicator vectors of the components, and $\operatorname{rank}C=v-c$. For the equivalences: $c=1$ iff $\operatorname{rank}C=v-1$; and since $C$ is symmetric, $\mathcal C(C)$ is the orthogonal complement of its null space, which equals $\mathbf 1^{\perp}=\{\lambda:\sum\lambda_i=0\}$ iff the null space is spanned by $\mathbf 1$, i.e. iff $c=1$. Apply Theorem 8.29(4). ∎

The formula says that $C$ is the Laplacian of a weighted graph on the treatments, in which $i$ and $i'$ are joined with weight $\sum_jn_{ij}n_{i'j}/k_j$. Statisticians and electrical engineers meet here: $\operatorname{Var}(\hat\tau_i-\hat\tau_{i'})/\sigma^2=(e_i-e_{i'})^{\mathsf T}C^-(e_i-e_{i'})$ is exactly the effective resistance between $i$ and $i'$ when each weight is read as a conductance. Treatments that meet often, directly or through many paths, are compared precisely; Example 1 makes this concrete.

📘 Definition 8.17 — Orthogonal block design
A block design is orthogonal if its incidence matrix has proportional frequencies: $$n_{ij}=\frac{r_ik_j}{n}\quad\text{for all }i,j,\qquad\text{i.e.}\quad N=\frac{rk^{\mathsf T}}{n}.$$
📐 Theorem 8.31 — Orthogonal designs
  1. The design is orthogonal iff the treatment and block indicator spaces, each with the constant removed, are orthogonal in $\mathbb R^n$.
  2. An orthogonal design is connected, with $C=R-rr^{\mathsf T}/n$ and $Q_i=T_i-r_iG/n$; the BLUE of $\tau_i-\tau_{i'}$ is $\bar y_{i\cdot}-\bar y_{i'\cdot}$ (unadjusted means), and the adjusted treatment sum of squares equals the unadjusted one, $\sum_iT_i^2/r_i-G^2/n$.
  3. Conversely, if $Q_i=T_i-r_iG/n$ for all data $y$ (adjustment for blocks never changes anything), the design is orthogonal.
  4. Every RBD is orthogonal; an incomplete design is never orthogonal.

Proof. (1) Let $x_i$ and $z_j$ be the indicator vectors of treatment $i$ and block $j$ in $\mathbb R^n$. Their centred versions have inner product $$\Bigl\langle x_i-\tfrac{r_i}n\mathbf 1,\ z_j-\tfrac{k_j}n\mathbf 1\Bigr\rangle=n_{ij}-\frac{r_ik_j}n-\frac{r_ik_j}n+\frac{r_ik_j}{n}=n_{ij}-\frac{r_ik_j}{n}.$$ (2) If $N=rk^{\mathsf T}/n$ then, since $k^{\mathsf T}K^{-1}k=\sum_jk_j=n$, $$NK^{-1}N^{\mathsf T}=\frac1{n^2}r\,(k^{\mathsf T}K^{-1}k)\,r^{\mathsf T}=\frac{rr^{\mathsf T}}{n},$$ and $NK^{-1}B=\frac{r}{n}k^{\mathsf T}K^{-1}B=\frac rn\sum_jB_j=\frac{rG}{n}$. Every $n_{ij}=r_ik_j/n\gt0$, so all treatments share every block and the design is connected. One solution of $C\hat\tau=Q$ is $\hat\tau_i=T_i/r_i-G/n$ (check: $R\hat\tau=T-rG/n$ and $r^{\mathsf T}\hat\tau=G-G=0$), so $\hat\tau_i-\hat\tau_{i'}=\bar y_{i\cdot}-\bar y_{i'\cdot}$ and $$\hat\tau^{\mathsf T}Q=\sum_i\Bigl(\frac{T_i}{r_i}-\frac Gn\Bigr)\Bigl(T_i-\frac{r_iG}{n}\Bigr)=\sum_i\frac{T_i^2}{r_i}-\frac{G^2}{n}.$$ (3) $Q=T-NK^{-1}B$ and $T-rG/n$ are linear in $y$; equality for all $y$ means $NK^{-1}X_\beta^{\mathsf T}=\frac rn\mathbf 1^{\mathsf T}$. Applying both sides to an observation in block $j$ gives $n_{ij}/k_j=r_i/n$. (4) For the RBD, $N=J$, $r=b\mathbf 1$, $k=v\mathbf 1$, $n=bv$: $rk^{\mathsf T}/n=J$. If some $n_{ij}=0$, then $n_{ij}\ne r_ik_j/n\gt0$. ∎

A complete design need not be orthogonal: two blocks of sizes $4$ and $3$ containing treatment 1 twice in the first block and once in the second (and treatments 2, 3 once in each) have $r=(3,2,2)$, $k=(4,3)$, $n=7$, and $n_{11}=2\ne\frac{3\cdot4}{7}$. "Complete", "connected" and "orthogonal" are three different properties.

✏️ Example 1 — A chain of blocks, and a design that falls apart
(a) Four treatments are tested in three blocks of size two: $\{1,2\}$, $\{2,3\}$, $\{3,4\}$. Find $N$, $C$ and the variances of $\hat\tau_1-\hat\tau_2$, $\hat\tau_1-\hat\tau_3$, $\hat\tau_1-\hat\tau_4$. (b) In the design with blocks $\{1,2\}$, $\{1,2\}$, $\{3,4\}$, $\{3,4\}$, which contrasts are estimable?
(a) Matrices $$N=\begin{pmatrix}1&0&0\\1&1&0\\0&1&1\\0&0&1\end{pmatrix},\quad r=(1,2,2,1),\quad k=(2,2,2).$$ $$C=\begin{pmatrix}\tfrac12&-\tfrac12&0&0\\-\tfrac12&1&-\tfrac12&0\\0&-\tfrac12&1&-\tfrac12\\0&0&-\tfrac12&\tfrac12\end{pmatrix},$$ the Laplacian of the path $1-2-3-4$ with weight $\frac12$ on each edge. Rank $3$: connected.
(a) Variances Treatments 1 and 2 meet only in block 1, so $\hat\tau_1-\hat\tau_2=y_{11}-y_{21}$ with variance $2\sigma^2$. Then $\hat\tau_1-\hat\tau_3=(y_{11}-y_{21})+(y_{22}-y_{32})$: two independent block differences, variance $4\sigma^2$; and $\operatorname{Var}(\hat\tau_1-\hat\tau_4)=6\sigma^2$. In resistance language: three resistors of $2$ ohms in series. (A generalised inverse of $C$ gives the same numbers.)
(b) $r=(2,2,2,2)$, $k=(2,2,2,2)$, and $C=\begin{pmatrix}1&-1\\-1&1\end{pmatrix}\oplus\begin{pmatrix}1&-1\\-1&1\end{pmatrix}$ has rank $2=v-2$: two components $\{1,2\}$ and $\{3,4\}$. Estimable: $\lambda^{\mathsf T}\tau$ with $\lambda_1+\lambda_2=0$ and $\lambda_3+\lambda_4=0$, i.e. combinations of $\tau_1-\tau_2$ and $\tau_3-\tau_4$ (each with variance $\sigma^2$, from two block differences). $\tau_1-\tau_3$ is not estimable: any difference between the groups is indistinguishable from a difference between the blocks that contain them.

Balanced Incomplete Block Designs

📘 Definition 8.18 — Balanced incomplete block design
A balanced incomplete block design BIBD$(v,b,r,k,\lambda)$ is an arrangement of $v$ treatments in $b$ blocks such that every block contains $k\lt v$ distinct treatments, every treatment occurs in $r$ blocks, and every pair of distinct treatments occurs together in exactly $\lambda$ blocks. It is symmetric if $b=v$. The efficiency factor is $E=\dfrac{\lambda v}{rk}$.
📐 Theorem 8.32 — Parameter relations and Fisher's inequality

For a BIBD$(v,b,r,k,\lambda)$:

  1. $bk=vr$ and $\lambda(v-1)=r(k-1)$; hence $r\gt\lambda$;
  2. $NN^{\mathsf T}=(r-\lambda)I_v+\lambda J_v$, with eigenvalues $rk$ (once) and $r-\lambda$ ($v-1$ times), so $\det NN^{\mathsf T}=rk(r-\lambda)^{v-1}\gt0$;
  3. (Fisher's inequality) $b\ge v$, and therefore $r\ge k$.

Proof. (1) Count the incident pairs (treatment, block containing it) in two ways: each of $v$ treatments lies in $r$ blocks, and each of $b$ blocks contains $k$ treatments, so $vr=bk$. Next fix a treatment $i$ and count pairs $(i',j)$ with $i'\ne i$ and both $i,i'$ in block $j$: each of the $r$ blocks containing $i$ contributes $k-1$ such $i'$, while each of the $v-1$ treatments $i'\ne i$ meets $i$ in $\lambda$ blocks; so $r(k-1)=\lambda(v-1)$. Since $k\lt v$, $\lambda(v-1)=r(k-1)\lt r(v-1)$, so $\lambda\lt r$. (2) $(NN^{\mathsf T})_{ii'}=\sum_jn_{ij}n_{i'j}$ is the number of blocks containing both $i$ and $i'$: $r$ on the diagonal (binary design) and $\lambda$ off it. $J$ has eigenvalue $v$ on $\mathbf 1$ and $0$ on $\mathbf 1^\perp$, so $NN^{\mathsf T}$ has eigenvalues $r-\lambda+\lambda v=r+r(k-1)=rk$ and $r-\lambda$. (3) By (1)–(2), $NN^{\mathsf T}$ is a nonsingular $v\times v$ matrix, so $v=\operatorname{rank}NN^{\mathsf T}\le\operatorname{rank}N\le b$. Then $r=bk/v\ge k$. ∎

The relations are necessary, not sufficient: $(v,k,\lambda)$ determines $r=\lambda(v-1)/(k-1)$ and $b=vr/k$, which must be integers with $b\ge v$; even then a design may fail to exist (Theorem 8.34 gives one more obstruction). Useful consequences: $b-v=\frac vk(r-k)$, so $b\ge v+r-k$; and $E=\frac{\lambda v}{rk}=\frac{v(k-1)}{k(v-1)}\lt1$.

📐 Theorem 8.33 — Intrablock analysis of a BIBD

For a BIBD$(v,b,r,k,\lambda)$:

  1. $C=\dfrac{\lambda v}{k}\Bigl(I-\dfrac1vJ\Bigr)$; the design is connected;
  2. with $Q_i=T_i-\frac1k\sum_{j\ni i}B_j$, the estimates with $\sum\hat\tau_i=0$ are $\hat\tau_i=\dfrac{k}{\lambda v}Q_i$, and every contrast $\sum c_i\tau_i$ has BLUE $\frac k{\lambda v}\sum c_iQ_i$ with variance $\frac{k\sigma^2}{\lambda v}\sum c_i^2$; in particular $$\operatorname{Var}(\hat\tau_i-\hat\tau_{i'})=\frac{2k\sigma^2}{\lambda v}\quad\text{for every pair};$$
  3. the adjusted treatment sum of squares is $\dfrac{k}{\lambda v}\sum_iQ_i^2$ on $v-1$ df, and the ANOVA table is: blocks (unadjusted) $b-1$ df, treatments (adjusted) $v-1$, error $bk-b-v+1$, total $bk-1$;
  4. relative to an RBD with the same replication $r$ and the same $\sigma^2$, the variance of an elementary contrast is larger by the factor $1/E$, where $E=\frac{\lambda v}{rk}\lt1$.

Proof. (1) $R=rI$ and $K=kI$, so by Theorem 8.32(2) $$C=rI-\frac1kNN^{\mathsf T}=\frac{(r(k-1)+\lambda)I-\lambda J}{k}=\frac{\lambda(v-1)+\lambda}{k}I-\frac{\lambda}kJ=\frac{\lambda v}{k}\Bigl(I-\frac1vJ\Bigr),$$ using $r(k-1)=\lambda(v-1)$. It has rank $v-1$, so the design is connected (Theorem 8.30). (2) $\sum_iQ_i=\mathbf 1^{\mathsf T}Q=0$ (since $\mathbf 1^{\mathsf T}C=0$ and Theorem 8.29), so with $\sum\hat\tau_i=0$ the equations $C\hat\tau=Q$ read $\frac{\lambda v}{k}\hat\tau=Q$. For a contrast, $\frac{k}{\lambda v}I$ acts as a generalised inverse of $C$ on $\mathbf 1^\perp$: $c^{\mathsf T}\frac k{\lambda v}C\frac{k}{\lambda v}c=\frac{k}{\lambda v}c^{\mathsf T}c$, and Theorem 8.29(2),(4) give the variance. (3) $\hat\tau^{\mathsf T}Q=\frac k{\lambda v}\sum Q_i^2$; the error df are $n-b-(v-1)$ with $n=bk$. (4) In an RBD with $r$ replicates, $\operatorname{Var}(\bar y_{i\cdot}-\bar y_{i'\cdot})=2\sigma^2/r$, and $\frac{2\sigma^2/r}{2k\sigma^2/(\lambda v)}=\frac{\lambda v}{rk}=E$. ∎

The efficiency factor compares designs with the same $\sigma^2$; in practice the point of small blocks is that their $\sigma^2$ is smaller, which can more than compensate for $E\lt1$. The intrablock analysis ignores the information in the block totals; when blocks are random, that interblock information can be recovered and combined with the intrablock estimates (Yates, 1940), but the intrablock analysis is what is examined.

✏️ Example 2 — Which parameter sets can be BIBDs?
Decide whether a BIBD exists with (a) $v=6$, $k=3$, $\lambda=2$; (b) $v=8$, $k=3$, $\lambda=1$; (c) $v=16$, $k=6$, $\lambda=1$; (d) $v=7$, $k=3$, $\lambda=1$; (e) $v=22$, $k=7$, $\lambda=2$.
(a) Exists $r=\frac{2\cdot5}{2}=5$, $b=\frac{6\cdot5}{3}=10\ge6$. A design: $\{1,2,3\},\{1,2,4\},\{1,3,5\},\{1,4,6\},\{1,5,6\},\{2,3,6\},\{2,4,5\},\{2,5,6\},\{3,4,5\},\{3,4,6\}$ — each of the $15$ pairs occurs in exactly two blocks (check, e.g., $\{1,2\}$: blocks 1, 2; $\{5,6\}$: blocks 5, 8).
(b) No $r=\frac{1\cdot7}{2}$ is not an integer.
(c) No $r=\frac{15}{5}=3$ and $b=\frac{16\cdot3}{6}=8\lt16$, violating Fisher's inequality, although both are integers.
(d) Exists $r=3$, $b=7$: the Fano plane, the translates of $\{1,2,4\}$ modulo $7$ (Theorem 8.35(3)).
(e) No $r=\frac{2\cdot21}{6}=7$, $b=22=v$: a symmetric design with $v$ even, so $k-\lambda=5$ would have to be a perfect square (Theorem 8.34). It is not.
✏️ Example 3 — Intrablock analysis of a BIBD
Four treatments are compared in four blocks of three plots, each block omitting one treatment ($v=b=4$, $r=k=3$, $\lambda=2$). Yields: block 1 $\{1,2,3\}$: $14,17,12$; block 2 $\{1,2,4\}$: $16,21,20$; block 3 $\{1,3,4\}$: $12,10,16$; block 4 $\{2,3,4\}$: $18,15,22$ (listed in treatment order). Carry out the intrablock analysis.
Totals Blocks $B=(43,57,38,55)$, $G=193$. Treatments $T=(42,56,37,58)$; unadjusted means $14,\ 18.67,\ 12.33,\ 19.33$.
Adjusted totals $Q_1=42-\frac{43+57+38}{3}=-4$; $Q_2=56-\frac{43+57+55}3=\frac{13}3$; $Q_3=37-\frac{43+38+55}{3}=-\frac{25}3$; $Q_4=58-\frac{57+38+55}3=8$. (Check: $\sum Q_i=0$.)
Estimates $\frac{k}{\lambda v}=\frac38$: $\hat\tau=(-1.5,\ 1.625,\ -3.125,\ 3.0)$. Adjusted means $\bar y+\hat\tau_i$ with $\bar y=16.08$: $14.58,\ 17.71,\ 12.96,\ 19.08$ — treatment 1 had been unlucky (it missed the best block 4), treatment 4 lucky.
ANOVA $\text{CF}=193^2/12=3104.08$, $\sum y^2=3259$, $\text{SST}=154.92$. Blocks (unadj.): $\frac{43^2+57^2+38^2+55^2}{3}-\text{CF}=84.92$ on $3$ df. Treatments (adj.): $\frac38\left(16+\frac{169}9+\frac{625}9+64\right)=63.08$ on $3$ df. Error: $154.92-84.92-63.08=6.92$ on $12-4-4+1=5$ df, $\text{MSE}=1.383$. $F=\frac{63.08/3}{1.383}=15.2\gt F_{3,5,0.05}=5.41$ ($p=0.006$).
Precision $\operatorname{Var}(\hat\tau_i-\hat\tau_{i'})=\frac{2\cdot3}{2\cdot4}\sigma^2=0.75\sigma^2$, $\text{SE}=\sqrt{0.75\times1.383}=1.02$, for every pair. $E=\frac{2\cdot4}{3\cdot3}=\frac89$. (A least squares fit with block and treatment effects reproduces these numbers exactly.)
📐 Theorem 8.34 — Symmetric BIBDs

In a symmetric BIBD ($b=v$, hence $r=k$):

  1. $N^{\mathsf T}N=(k-\lambda)I+\lambda J$: any two blocks have exactly $\lambda$ treatments in common; also $\lambda(v-1)=k(k-1)$;
  2. $(\det N)^2=k^2(k-\lambda)^{v-1}$; hence if $v$ is even, $k-\lambda$ is a perfect square (Schützenberger; part of the Bruck–Ryser–Chowla theorem).

Proof. (1) $N$ is square and $NN^{\mathsf T}=(k-\lambda)I+\lambda J$ is nonsingular (Theorem 8.32), so $N$ is invertible. Row sums of $N$ are $r=k$ and column sums are $k$: $NJ=kJ$ and $JN=kJ$. From $NJ=kJ$ we get $N^{-1}J=\frac1kJ$. Then $$N^{\mathsf T}=N^{-1}(NN^{\mathsf T})=N^{-1}\bigl((k-\lambda)I+\lambda J\bigr)=(k-\lambda)N^{-1}+\tfrac{\lambda}{k}J,$$ and $N^{\mathsf T}N=(k-\lambda)I+\frac\lambda kJN=(k-\lambda)I+\lambda J$. The $(j,j')$ entry of $N^{\mathsf T}N$ counts the treatments common to blocks $j$ and $j'$. (2) By Theorem 8.32(2) with $r=k$, $$(\det N)^2=\det(NN^{\mathsf T})=k^2(k-\lambda)^{v-1}.$$ Since $\det N$ is an integer, $(k-\lambda)^{v-1}=(\det N/k)^2$ is the square of a rational number, hence of an integer. If $v$ is even, $v-2$ is even, so $(k-\lambda)^{v-2}$ is a square and $k-\lambda=(k-\lambda)^{v-1}/(k-\lambda)^{v-2}$ is the square of a rational, hence of an integer. ∎

📐 Theorem 8.35 — Three constructions
  1. (Unreduced design.) All $\binom vk$ subsets of size $k$ form a BIBD with $b=\binom vk$, $r=\binom{v-1}{k-1}$, $\lambda=\binom{v-2}{k-2}$.
  2. (Complement.) If $v-k\ge2$, replacing every block of a BIBD$(v,b,r,k,\lambda)$ by its complement gives a BIBD$(v,\,b,\,b-r,\,v-k,\,b-2r+\lambda)$.
  3. (Difference sets.) If $D=\{d_1,\dots,d_k\}\subset\mathbb Z_v$ and every nonzero element of $\mathbb Z_v$ occurs exactly $\lambda$ times among the differences $d_s-d_t$ ($s\ne t$), then the $v$ translates $D+x$ ($x\in\mathbb Z_v$) form a symmetric BIBD$(v,v,k,k,\lambda)$.

Proof. (1) A treatment lies in the subsets that contain it and $k-1$ of the other $v-1$; a pair lies in those containing it and $k-2$ of the other $v-2$. (2) A treatment is absent from $b-r$ blocks. Two treatments are both absent from a block iff the block contains neither; by inclusion–exclusion the number of blocks containing at least one of them is $2r-\lambda$, so both are absent from $b-2r+\lambda$ blocks. The new block size $v-k$ is less than $v$, and is at least $2$ so that pairs make sense. (3) A treatment $i$ lies in $D+x$ iff $i-x\in D$, i.e. for the $k$ values $x=i-d_s$. Two treatments $i\ne i'$ both lie in $D+x$ iff $i-x=d_s$ and $i'-x=d_t$ for some $s,t$, which forces $d_s-d_t=i-i'$; each of the $\lambda$ pairs $(s,t)$ with that difference gives exactly one $x=i-d_s$, and distinct pairs give distinct $x$ (as $s$ determines $x$). ∎

✏️ Example 4 — A projective plane of order 3
Show that $D=\{0,1,3,9\}$ is a difference set modulo $13$, list the resulting design, verify that any two blocks meet in exactly one treatment, and find its efficiency factor.
Differences The $12$ ordered differences are $\pm1,\pm3,\pm9,\pm2\ (3-1),\pm8\ (9-1),\pm6\ (9-3)$, i.e. $1,12,3,10,9,4,2,11,8,5,6,7$ modulo $13$: every nonzero residue exactly once. So $\lambda=1$, and with $k=4$: $\lambda(v-1)=12=k(k-1)$ ✓.
Blocks $D+x$ for $x=0,\dots,12$: $\{0,1,3,9\},\{1,2,4,10\},\{2,3,5,11\},\{3,4,6,12\},\{4,5,7,0\},\{5,6,8,1\},\{6,7,9,2\},\{7,8,10,3\},\{8,9,11,4\},\{9,10,12,5\},\{10,11,0,6\},\{11,12,1,7\},\{12,0,2,8\}$: a BIBD$(13,13,4,4,1)$.
Blocks meet once Theorem 8.34(1) with $\lambda=1$; directly, $\{0,1,3,9\}\cap\{1,2,4,10\}=\{1\}$, $\{0,1,3,9\}\cap\{4,5,7,0\}=\{0\}$, and so on. Treatments as points and blocks as lines, this is the projective plane of order $3$: two points lie on one line, two lines meet in one point.
Efficiency $E=\frac{\lambda v}{rk}=\frac{13}{16}=0.8125$; every elementary contrast has variance $\frac{2k\sigma^2}{\lambda v}=\frac{8}{13}\sigma^2$. Since $v=13$ is odd, the obstruction of Theorem 8.34(2) does not apply (and $k-\lambda=3$ is not a square — no contradiction).
Interactive 1 — Building a Block Design: Incidence Matrix and Concordance Graph

Left: the incidence matrix $N$ (treatments × blocks). Right: the treatments on a circle; the line between $i$ and $i'$ gets thicker each time a block containing both is added — its final thickness is $\lambda_{ii'}=(NN^{\mathsf T})_{ii'}$. Press Play to add the blocks one at a time. A BIBD ends with all lines equally thick, so every contrast $\hat\tau_i-\hat\tau_{i'}$ has the same variance; an unbalanced design gives unequal variances; a disconnected design leaves treatments that can never be compared.

v · b · r · k—
pair counts λ (min–max)—
rank C · connected?—
variance of a pairwise contrast / σ² (min–max)—

2ᵏ Factorial Experiments

In a factorial experiment the treatments are all combinations of the levels of several factors. Studying the factors together rather than one at a time has two advantages: every observation is used for every main effect (so the same precision is reached with far fewer runs), and interactions — the effect of one factor depending on the level of another — can be detected at all. The simplest and most important case has $k$ factors, each at two levels.

📘 Definition 8.19 — The $2^k$ factorial and its effects

Factors $A,B,C,\dots$ have a low and a high level, coded $0$ and $1$. A treatment combination is $t=(t_A,t_B,\dots)\in\{0,1\}^k$, written as the product of the letters at their high level: $(1)$ (all low), $a$, $b$, $ab$, $c$, $ac$, $bc$, $abc$, … — this is the standard order, in which $t$ has index $t_A+2t_B+4t_C+\cdots$. The same symbol also denotes the total of the $r$ observations on that combination. For a nonempty set of letters $X$ (an effect: a main effect $A$, a two-factor interaction $AB$, …) define the sign $$s_X(t)=\prod_{\ell\in X}(2t_\ell-1)\in\{-1,+1\},$$ the contrast $[X]=\sum_ts_X(t)\,(t)$ (a signed sum of the $2^k$ totals), the effect estimate and its sum of squares $$\hat X=\frac{[X]}{r\,2^{k-1}},\qquad\text{SS}_X=\frac{[X]^2}{r\,2^k}.$$

Formally, $[X]$ is obtained by expanding the product $\prod_{\ell}(\ell\pm1)$ with "$-$" for the letters in $X$ and "$+$" for the others, e.g. in a $2^2$: $[A]=(a-1)(b+1)=ab+a-b-(1)$, $[B]=ab-a+b-(1)$, $[AB]=(a-1)(b-1)=ab-a-b+(1)$.

With these scalings, $\hat A$ is (mean response at high $A$) − (mean at low $A$), averaged over all levels of the other factors; $\widehat{AB}$ is half the difference between the effect of $A$ at high $B$ and at low $B$; $\widehat{ABC}$ is half the difference between the $AB$ interaction at high $C$ and at low $C$; and so on.

📐 Theorem 8.36 — Factorial effects are orthogonal contrasts

In a $2^k$ factorial with $r$ replicates (in a CRD or in complete blocks) and model errors iid $N(0,\sigma^2)$:

  1. each $s_X$ ($X\ne\varnothing$) is a contrast ($\sum_ts_X(t)=0$), and $\sum_ts_X(t)s_Y(t)=0$ for $X\ne Y$: the $2^k-1$ effects are mutually orthogonal;
  2. $E[\hat X]$ is the true effect, $\operatorname{Var}(\hat X)=\dfrac{\sigma^2}{r\,2^{k-2}}$, and the estimates are independent;
  3. $\sum_X\text{SS}_X$ equals the treatment sum of squares $\sum_t(t)^2/r-G^2/(r2^k)$; each $\text{SS}_X$ has $1$ df, and $\text{SS}_X/\sigma^2\sim\chi^2_1$ when the effect is zero.

Proof. (1) Since $(2t_\ell-1)^2=1$, $s_X(t)s_Y(t)=s_{X\triangle Y}(t)$, where $X\triangle Y$ is the symmetric difference (letters in exactly one of $X,Y$). So it suffices that $\sum_ts_Z(t)=0$ for $Z\ne\varnothing$: pick $\ell\in Z$ and pair each $t$ with the combination $t'$ that differs from it only in factor $\ell$; then $s_Z(t')=-s_Z(t)$. (2) $[X]$ is a signed sum of $r2^k$ independent observations, so $\operatorname{Var}([X])=r2^k\sigma^2$ and $\operatorname{Var}(\hat X)=\frac{r2^k\sigma^2}{r^22^{2k-2}}=\frac{\sigma^2}{r2^{k-2}}$. Orthogonal contrasts of independent normal totals are independent. (3) The vectors $s_X/\sqrt{2^k}$, together with $\mathbf 1/\sqrt{2^k}$, form an orthonormal basis of $\mathbb R^{2^k}$ by (1). Expanding the vector of treatment means $\bar y_t=(t)/r$ in this basis and using Parseval's identity, $r\sum_t(\bar y_t-\bar y)^2=r\sum_X\frac{(\sum_ts_X(t)\bar y_t)^2}{2^k}=\sum_X\frac{[X]^2}{r2^k}$. The distributional statement is §6.1 for a single contrast. ∎

📐 Theorem 8.37 — Yates' algorithm
Write the $2^k$ treatment totals in standard order as a column. Form a new column whose first half consists of the sums of successive pairs ($1{+}2$, $3{+}4$, …) and whose second half consists of the differences of the same pairs, second minus first ($2{-}1$, $4{-}3$, …). After $k$ such passes, the entry in the row of treatment combination $e$ is the contrast $[X]$ of the effect $X$ with the same letters as $e$ (the first entry is the grand total $G$).

Proof. Identify a position with its index written in binary, $t=(t_1,\dots,t_k)$, $t_1$ the least significant bit (factor $A$). One pass combines positions $2p$ and $2p+1$, which differ only in $t_1$, and writes the result at position $p$ (sum) or $2^{k-1}+p$ (difference). In binary, the output index is $(t_2,\dots,t_k,e_1)$, where $e_1=0$ for a sum and $1$ for a difference: the bits rotate one place, and the processed bit is replaced by the choice "sum or difference", with coefficient $c(0,t_1)=1$ for a sum and $c(1,t_1)=2t_1-1$ for a difference. By induction, after $m$ passes the entry at position $(t_{m+1},\dots,t_k,e_1,\dots,e_m)$ is $\sum_{t_1,\dots,t_m}\prod_{\ell\le m}c(e_\ell,t_\ell)\,(t)$. After $k$ passes the bits have rotated back, the entry at position $(e_1,\dots,e_k)$ is $\sum_t\prod_\ell c(e_\ell,t_\ell)(t)$, and $\prod_\ell c(e_\ell,t_\ell)=\prod_{\ell:e_\ell=1}(2t_\ell-1)=s_X(t)$ for $X=\{\ell:e_\ell=1\}$. ∎

✏️ Example 5 — A $2^3$ experiment by Yates' algorithm
An engineer studies the strength of a weld with factors $A$ (current), $B$ (electrode type), $C$ (preheating), each at two levels, in two complete replicates (blocks). Replicate I, in standard order: $20,28,23,37,21,30,22,38$; replicate II: $22,30,21,35,19,31,24,36$. Estimate all effects and test them.
Yates Totals $(1)\dots abc$: $42,58,44,72,40,61,46,74$.
comb.total(1)(2)(3)effectestimate $=(3)/8$SS $=(3)^2/16$
(1)42100216437$G$——
a5811622193$A$11.625540.56
b441014435$B$4.37576.56
ab721204919$AB$2.37522.56
c4016165$C$0.6251.56
ac6128195$AC$0.6251.56
bc4621123$BC$0.3750.56
abc74287−5$ABC$−0.6251.56
ANOVA $\sum y^2=12595$, $\text{CF}=437^2/16=11935.56$, $\text{SST}=659.44$; replicates (blocks): $\frac{219^2+218^2}8-\text{CF}=0.06$ on $1$ df; treatments $644.94$ on $7$ df (the sum of the seven SS); error $659.44-0.06-644.94=14.44$ on $7$ df, $\text{MSE}=2.063$.
Tests $F_{1,7,0.05}=5.59$. $A$: $F=262$; $B$: $F=37.1$; $AB$: $F=10.9$ ($p=0.013$); $C$, $AC$, $BC$, $ABC$: $F\le0.76$. Current, electrode and their interaction matter; preheating does not.
Interpretation The effect of current is $\frac{(58-42)+(61-40)}{4}=9.25$ with the low-level electrode and $\frac{(72-44)+(74-46)}4=14$ with the high-level one; $\widehat{AB}=\frac{14-9.25}{2}=2.375$ ✓. Report the $A$ effect separately at each level of $B$, not only the average $11.6$. Each estimate has standard error $\sqrt{2.063/4}=0.72$ (Theorem 8.36(2) with $r=2$, $k=3$).

Confounding: Factorials in Incomplete Blocks

A $2^k$ factorial needs blocks of $2^k$ plots to be run in complete blocks, which quickly becomes impossible ($2^5=32$). The way out is to give up the information on one or more effects, chosen to be ones we expect to be negligible (usually high-order interactions), by making their contrasts coincide with differences between blocks.

📘 Definition 8.20 — Confounding

A replicate of a $2^k$ factorial is arranged in $2^p$ blocks of $2^{k-p}$ plots. An effect $X$ is confounded with blocks if its sign $s_X(t)$ is constant on every block: then its contrast is a contrast between block totals and cannot be separated from block differences. To construct the blocks, choose $p$ independent effects $X_1,\dots,X_p$ (the defining contrasts), and put $t$ into the block labelled $(L_1(t),\dots,L_p(t))$, where $$L_X(t)=\sum_{\ell\in X}t_\ell\pmod 2.$$ The block containing $(1)$ is the principal block. The generalised interaction of $X$ and $Y$ is $XY:=X\triangle Y$ (multiply and delete squared letters: $ABC\cdot BCD=AB^2C^2D=AD$). If the same effects are confounded in all replicates the confounding is total; if different effects are confounded in different replicates it is partial.

📐 Theorem 8.38 — Construction of blocks and the confounded effects

Identify treatment combinations and effects with vectors in $\mathbb F_2^{\,k}$ ($X\leftrightarrow e_X$, its indicator), so that $L_X(t)=e_X\cdot t$ and $s_X(t)=(-1)^{|X|}(-1)^{e_X\cdot t}$. With defining contrasts $X_1,\dots,X_p$ (linearly independent over $\mathbb F_2$):

  1. the principal block $H=\{t:L_{X_1}(t)=\cdots=L_{X_p}(t)=0\}$ is a subgroup (subspace) of order $2^{k-p}$, closed under multiplication of treatment symbols with squares deleted ($ab\cdot ac=bc$); the other blocks are its cosets, obtained by multiplying the principal block by any combination not in it;
  2. the effects confounded with blocks are exactly the $2^p-1$ nonzero elements of the span of $e_{X_1},\dots,e_{X_p}$: the defining contrasts and all their generalised interactions;
  3. every other effect has a contrast that sums to zero within each block: it is orthogonal to blocks and estimated, from within blocks, with full efficiency.

Proof. $2t_\ell-1=-(-1)^{t_\ell}$, so $s_X(t)=\prod_{\ell\in X}\bigl(-(-1)^{t_\ell}\bigr)=(-1)^{|X|}(-1)^{\sum_{\ell\in X}t_\ell}$. (1) $H$ is the kernel of the linear map $t\mapsto(e_{X_1}\cdot t,\dots,e_{X_p}\cdot t)$ from $\mathbb F_2^{\,k}$ onto $\mathbb F_2^{\,p}$ (onto, because the $e_{X_i}$ are independent), so it is a subspace of dimension $k-p$, and the blocks — the fibres of the map — are its cosets. Adding vectors mod $2$ is multiplying symbols with squares deleted. (2), (3) Let $S=\operatorname{span}\{e_{X_i}\}$. If $e_X\in S$, say $e_X=\sum_{i\in I}e_{X_i}$, then $e_X\cdot t=\sum_{i\in I}L_{X_i}(t)$ is constant on each block, so $X$ is confounded. If $e_X\notin S$: the orthogonal complement of $H$ is $S$ (dimensions $p$ and $k-p$, and $S\subseteq H^{\perp}$), so there is $h\in H$ with $e_X\cdot h=1$. Then $t\mapsto t+h$ maps each block (coset $t_0+H$) onto itself and changes $e_X\cdot t$, hence flips the sign $s_X$. So $s_X$ takes the values $\pm1$ equally often in every block: the contrast is orthogonal to every block indicator, block effects cancel from it, and its variance is the same as without blocking. ∎

📐 Theorem 8.39 — Partial confounding
Suppose a $2^k$ factorial is run in $r$ replicates, each in blocks, and the effect $X$ is confounded in $r_c$ of them ($0\le r_c\lt r$). Then the intrablock estimate of $X$ uses only the $r-r_c$ replicates in which it is unconfounded: $\hat X=[X]_{\text{unconf}}/\bigl((r-r_c)2^{k-1}\bigr)$, with $\operatorname{Var}(\hat X)=\dfrac{\sigma^2}{(r-r_c)2^{k-2}}$ and $\text{SS}_X=[X]_{\text{unconf}}^2/\bigl((r-r_c)2^k\bigr)$ on $1$ df. The relative information on $X$, compared with no confounding, is $(r-r_c)/r$.

Proof. In a replicate where $X$ is confounded, its contrast is a contrast of block totals (Theorem 8.38(2)), so once block effects are eliminated it carries no information. In each replicate where it is unconfounded, its contrast is orthogonal to the blocks (Theorem 8.38(3)) and has expectation $2^{k-1}X$ and variance $2^k\sigma^2$, independently over replicates. Summing over the $r-r_c$ such replicates and applying Theorem 8.36 with $r-r_c$ in place of $r$ gives the formulas. The ratio of variances is $\frac{\sigma^2/(r2^{k-2})}{\sigma^2/((r-r_c)2^{k-2})}=\frac{r-r_c}{r}$. ∎

✏️ Example 6 — Constructing blocks; an unavoidable sacrifice
(a) Arrange a $2^3$ factorial in two blocks of four with $ABC$ confounded. (b) Arrange a $2^4$ factorial in four blocks of four confounding $ABC$ and $BCD$. Which other effect is confounded? (c) Prove that whenever a $2^4$ factorial is arranged in four blocks of four, some main effect or two-factor interaction is confounded.
(a) $L=t_A+t_B+t_C\pmod2$. Principal block ($L=0$, an even number of letters): $\{(1),ab,ac,bc\}$; other block: $\{a,b,c,abc\}$ (multiply the principal block by $a$). $ABC$ is the only confounded effect; each of the six others has two $+$ and two $-$ signs in each block.
(b) Blocks by $(L_{ABC},L_{BCD})$: $(0,0)$: $\{(1),bc,abd,acd\}$; $(1,0)$: $\{a,abc,bd,cd\}$; $(1,1)$: $\{b,c,ad,abcd\}$; $(0,1)$: $\{ab,ac,d,bcd\}$. Check the principal block is closed: $bc\cdot abd=ab^2cd=acd$ ✓. Also confounded: the generalised interaction $ABC\cdot BCD=AD$ — a two-factor interaction, the price of this choice.
(c) Proof The confounded effects are the three nonzero vectors $u,w,u+w$ of a $2$-dimensional subspace of $\mathbb F_2^{\,4}$. For each coordinate (letter) $\ell$, the number of these three vectors with a $1$ in position $\ell$ is $0$ or $2$ (if $u_\ell=1,w_\ell=0$ then $(u+w)_\ell=1$; if both are $1$ then $(u+w)_\ell=0$). So the total number of letters in the three effects is at most $2\cdot4=8$, and the shortest of them has at most $\lfloor8/3\rfloor=2$ letters. ∎ The best one can do is to confound one two-factor interaction and two three-factor ones (as in (b)), or e.g. $ABCD$ with two two-factor interactions — clearly worse.
Interactive 2 — Yates' Algorithm and the Confounding Cube

Left: Example 5's totals in standard order. Each pass replaces the column by sums of pairs (top half) and differences of pairs (bottom half); press Play or Next pass to watch the three passes produce every effect contrast. Right: the eight treatment combinations as corners of a cube, coloured by block for the chosen confounding scheme. Confounded effects (including generalised interactions) are struck out in the table: their contrasts have become block differences.

pass—
confounded effects—
principal block—
largest effects—
✏️ Example 7 — Partial confounding in a $2^3$ experiment
A $2^3$ factorial is run in four replicates, each in two blocks of four, with $ABC$ confounded in replicate I, $AB$ in II, $AC$ in III and $BC$ in IV. (a) Write the blocks of replicate II. (b) Give the ANOVA degrees of freedom. (c) Compare the variances of $\hat A$ and $\widehat{AB}$, and compare the design with confounding $ABC$ in all four replicates.
(a) $L_{AB}=t_A+t_B$: $\{(1),ab,c,abc\}$ and $\{a,b,ac,bc\}$.
(b) df $32$ plots, total $31$. Blocks: $8-1=7$ (this includes replicate differences and the four confounded contrasts). Treatments: $A,B,C$ (full information) and $AB,AC,BC,ABC$ (each from three replicates): $7$ df. Error: $31-7-7=17$.
(c) Variances Main effects: $\operatorname{Var}(\hat A)=\frac{\sigma^2}{4\cdot2}=\frac{\sigma^2}{8}$. Each interaction is unconfounded in $3$ replicates: $\operatorname{Var}=\frac{\sigma^2}{3\cdot2}=\frac{\sigma^2}{6}$; relative information $\frac34$ (Theorem 8.39). $\widehat{AB}$ is computed from the totals of replicates I, III and IV only.
Comparison With $ABC$ totally confounded, the six other effects have full information ($\sigma^2/8$) and $ABC$ cannot be estimated at all from within blocks; error df $31-7-6=18$. Partial confounding trades a little information on each interaction for some information on all of them — the right choice when no interaction can be assumed negligible in advance.
🧭 Problem-solving strategy — block designs and confounding
  1. BIBD parameters: from any three of $v,b,r,k,\lambda$ use $bk=vr$ and $\lambda(v-1)=r(k-1)$; reject if $r$ or $b$ is not an integer, if $b\lt v$ (Fisher), or if $b=v$, $v$ even and $k-\lambda$ is not a square.
  2. Intrablock analysis: $Q_i=T_i-\frac1k\sum_{j\ni i}B_j$, $\hat\tau_i=\frac{k}{\lambda v}Q_i$, SS(adj) $=\frac k{\lambda v}\sum Q_i^2$, error df $bk-b-v+1$, $\operatorname{Var}(\hat\tau_i-\hat\tau_{i'})=\frac{2k\sigma^2}{\lambda v}$, efficiency $E=\frac{\lambda v}{rk}$.
  3. General design: $C=R-NK^{-1}N^{\mathsf T}$; connected iff rank $v-1$ (follow the treatment–block chains); orthogonal iff $n_{ij}=r_ik_j/n$.
  4. Factorial effects: Yates' algorithm with $k$ passes; estimate $=$ contrast$/(r2^{k-1})$, SS $=$ contrast$^2/(r2^k)$.
  5. Confounding: blocks by the parities $L_X(t)$ of the defining contrasts; principal block = even parity for all; confounded set = defining contrasts and all generalised interactions (multiply, delete squares); keep main effects and, if possible, two-factor interactions out of it.
⚠️ Counterexamples and pitfalls
  1. Integer parameters do not guarantee a BIBD. $(v,k,\lambda)=(16,6,1)$ gives integers $r=3$, $b=8$, but violates Fisher's inequality; $(22,7,2)$ passes Fisher ($b=v$) but fails the symmetric-design condition ($k-\lambda=5$ is not a square).
  2. Two blocks meet in $\lambda$ treatments only in a symmetric BIBD. In the unreduced design with $v=4$, $k=2$, blocks $\{1,2\}$ and $\{3,4\}$ share nothing while $\{1,2\}$, $\{1,3\}$ share one.
  3. Complete ≠ orthogonal. A complete block design with unequal frequencies (e.g. treatment 1 twice in one block) violates $n_{ij}=r_ik_j/n$; adjusted and unadjusted treatment SS then differ.
  4. Connected ≠ balanced. The chain $\{1,2\},\{2,3\},\{3,4\}$ is connected, but the variances of $\hat\tau_1-\hat\tau_2$ and $\hat\tau_1-\hat\tau_4$ are $2\sigma^2$ and $6\sigma^2$.
  5. Disconnected designs hide contrasts. With blocks $\{1,2\},\{1,2\},\{3,4\},\{3,4\}$, $\tau_1-\tau_3$ is not estimable, however many observations are taken.
  6. Unadjusted means are biased in incomplete blocks. In Example 3 the raw means rank treatments 2 and 4 as $18.67$ vs $19.33$, the adjusted ones $17.71$ vs $19.08$; with stronger block effects the order can reverse.
  7. $E\lt1$ does not mean a BIBD is worse. The efficiency factor compares at equal $\sigma^2$; smaller blocks usually have smaller $\sigma^2$.
  8. Generalised interactions are confounded too. Confounding $AB$ and $AC$ in a $2^3$ also confounds $BC$; confounding $ABC$ and $BCD$ in a $2^4$ also confounds $AD$. Confounding $ABCD$ and $ABC$ would confound the main effect $D$.
  9. A $2^4$ in four blocks of four must lose a two-factor interaction (or worse), by the counting argument of Example 6(c).
  10. Yates' algorithm gives contrasts, not effects. Divide by $r2^{k-1}$ for the effect, and use contrast$^2/(r2^k)$ for the SS; the treatment totals must be in standard order.
  11. Partial confounding still costs. The partially confounded effects are estimated from the unconfounded replicates only: variance larger by $r/(r-r_c)$.

Exercises

I
Block designs and the C-matrix.
  • 1 Write the incidence matrix, $r$, $k$ and the C-matrix of the design with blocks $\{1,2,3\}$, $\{1,4\}$, $\{2,4\}$, $\{3,4\}$. Is it connected? Proper? Equireplicate?
  • 2 Prove that the row sums of every C-matrix are zero and that $\mathbf 1^{\mathsf T}Q=0$.
  • 3 Six treatments are arranged in blocks $\{1,2\},\{3,4\},\{5,6\},\{2,3\}$. Find $\operatorname{rank}C$ and describe the estimable treatment contrasts.
  • 4 Show that the C-matrix of an RBD with $v$ treatments and $b$ blocks is $b(I-\frac1vJ)$, and deduce $\operatorname{Var}(\hat\tau_i-\hat\tau_{i'})$.
  • 5 Decide whether the designs with incidence matrices $N_1=\begin{pmatrix}2&1\\2&1\\4&2\end{pmatrix}$ and $N_2=\begin{pmatrix}2&1\\1&1\\1&1\end{pmatrix}$ are orthogonal.
  • 6 Prove that a connected binary design satisfies $n-b\ge v-1$, i.e. $\sum_j(k_j-1)\ge v-1$. How many blocks of size $2$ are needed, at least, to connect $v$ treatments?
  • 7 For the design of Exercise I.1, find $\operatorname{Var}(\hat\tau_1-\hat\tau_2)$ and $\operatorname{Var}(\hat\tau_1-\hat\tau_4)$, and explain which is smaller and why.
  • 8 Show that $\operatorname{Cov}(Q)=\sigma^2C$ directly in the case of a BIBD, by computing $\operatorname{Var}(Q_i)$.
1 Rows = treatments, columns = blocks: $N=\begin{pmatrix}1&1&0&0\\1&0&1&0\\1&0&0&1\\0&1&1&1\end{pmatrix}$, $r=(2,2,2,3)$, $k=(3,2,2,2)$. $C$ has diagonal $\frac76,\frac76,\frac76,\frac32$; $c_{12}=c_{13}=c_{23}=-\frac13$ (from the block of size $3$); $c_{14}=c_{24}=c_{34}=-\frac12$. Connected (rank $3$); neither proper nor equireplicate.
2 $C\mathbf 1=R\mathbf 1-NK^{-1}N^{\mathsf T}\mathbf 1=r-NK^{-1}k=r-N\mathbf 1=0$. $\mathbf 1^{\mathsf T}Q=\mathbf 1^{\mathsf T}T-\mathbf 1^{\mathsf T}NK^{-1}B=G-k^{\mathsf T}K^{-1}B=G-\mathbf 1^{\mathsf T}B=0$.
3 Components $\{1,2,3,4\}$ (via $1$–$2$–$3$–$4$) and $\{5,6\}$: $\operatorname{rank}C=6-2=4$. Estimable: $\sum\lambda_i\tau_i$ with $\lambda_1+\lambda_2+\lambda_3+\lambda_4=0$ and $\lambda_5+\lambda_6=0$.
4 $N=J_{v\times b}$, $R=bI$, $K=vI$: $C=bI-\frac1vJ_{v\times b}J_{b\times v}=bI-\frac bvJ$. On contrasts $\frac1bI$ is a generalised inverse, so $\operatorname{Var}(\hat\tau_i-\hat\tau_{i'})=\frac{2\sigma^2}{b}$, as in §8.3.
5 $N_1$: $r=(3,3,6)$, $k=(8,4)$, $n=12$; $rk^{\mathsf T}/n=\begin{pmatrix}2&1\\2&1\\4&2\end{pmatrix}=N_1$: orthogonal. $N_2$: $r=(3,2,2)$, $k=(4,3)$, $n=7$; $r_1k_1/n=12/7\ne2$: complete but not orthogonal.
6 Start with $v$ isolated treatments ($v$ components) and add the blocks one at a time. A block with $k_j$ distinct treatments joins at most $k_j$ components into one, reducing the number of components by at most $k_j-1$. To reach one component, $\sum_j(k_j-1)\ge v-1$. Blocks of size $2$: at least $v-1$ (a spanning tree; the chain of Example 1 is one).
7 With $C^-$ (e.g. via $(C+\frac1vJ)^{-1}$): $\operatorname{Var}(\hat\tau_1-\hat\tau_2)=\frac43\sigma^2$, $\operatorname{Var}(\hat\tau_1-\hat\tau_4)=\frac{10}9\sigma^2$. Treatments 1 and 2 meet only in the block of size $3$ (weight $\frac13$), while 1 and 4 meet in a block of size $2$ (weight $\frac12$) and are also linked through the paths via 2 and 3; more conductance, less resistance.
8 $Q_i=T_i-\frac1k\sum_{j\ni i}B_j$. Write each $B_j$ ($j\ni i$) as the plot of $i$ plus the other $k-1$ plots: $Q_i=\frac{k-1}{k}T_i-\frac1k(\text{sum of the }r(k-1)\text{ other plots in }i\text{'s blocks})$, so $\operatorname{Var}(Q_i)=\sigma^2\left[r\frac{(k-1)^2}{k^2}+\frac{r(k-1)}{k^2}\right]=\sigma^2\frac{r(k-1)}{k}$. Theorem 8.33(1) gives $c_{ii}=\frac{\lambda v}{k}\cdot\frac{v-1}{v}=\frac{\lambda(v-1)}{k}=\frac{r(k-1)}{k}$ ✓.
II
Balanced incomplete block designs.
  • 1 A BIBD has $v=13$, $k=4$, $\lambda=1$. Find $r$ and $b$. Is it symmetric?
  • 2 Use the necessary conditions to decide which of these may exist: $(v,k,\lambda)=(11,5,2)$, $(10,4,1)$, $(15,7,3)$, $(6,3,1)$.
  • 3 Find the parameters of the complement of the Fano plane, and list its blocks.
  • 4 Find the parameters of the unreduced BIBD with $v=5$, $k=3$, and verify the two relations of Theorem 8.32.
  • 5 Prove that $r\gt\lambda$ in every BIBD, and that $b\ge v+r-k$.
  • 6 Show that $\{1,3,4,5,9\}$ (the nonzero squares modulo $11$) is a difference set in $\mathbb Z_{11}$ and give the parameters of the design it generates.
  • 7 In a BIBD with $v=4$, $k=3$, $\lambda=2$, the adjusted totals are $Q=(4,-2,1,-3)$. Find $\hat\tau$, the adjusted treatment SS, and $\operatorname{Var}(\hat\tau_1-\hat\tau_2)$ in terms of $\sigma^2$.
  • 8 Prove that no BIBD with $v=b=22$, $k=r=7$, $\lambda=2$ exists, while for $(16,16,6,6,2)$ Theorem 8.34 gives no obstruction. How many treatments do two blocks of the latter share?
  • 9 Show that $E=\frac{v(k-1)}{k(v-1)}$ and compute $E$ for the Fano plane and for the design of Example 3. For fixed $k$, what happens as $v\to\infty$?
1 $r=\frac{1\cdot12}{3}=4$, $b=\frac{13\cdot4}{4}=13$: symmetric (it is Example 4).
2 $(11,5,2)$: $r=5$, $b=11$, symmetric, $v$ odd — passes (and exists, Exercise II.6). $(10,4,1)$: $r=3$, $b=7.5$ — no. $(15,7,3)$: $r=7$, $b=15$, symmetric, $v$ odd — passes (it exists: the points and planes of a projective $3$-space over $\mathbb F_2$). $(6,3,1)$: $r=2.5$ — no.
3 $(7,7,7-3,7-3,7-6+1)=(7,7,4,4,2)$. Blocks: complements of the translates of $\{1,2,4\}$ mod $7$, i.e. the translates of $\{0,3,5,6\}$: $\{0,3,5,6\},\{1,4,6,0\},\{2,5,0,1\},\{3,6,1,2\},\{4,0,2,3\},\{5,1,3,4\},\{6,2,4,5\}$.
4 $b=\binom53=10$, $r=\binom42=6$, $\lambda=\binom31=3$. $bk=30=vr$; $\lambda(v-1)=12=r(k-1)$ ✓.
5 $\lambda(v-1)=r(k-1)$ with $k-1\lt v-1$ gives $\lambda\lt r$. Next, $b-v=\frac{vr}{k}-v=\frac vk(r-k)\ge r-k$ because $r\ge k$ (Fisher) and $v/k\gt1$.
6 Modulo $11$: $1=4-3=5-4$, $2=3-1=5-3$, $3=1-9=4-1$, $4=5-1=9-5$, $5=3-9=9-4$, and the remaining residues are the negatives of these ($6=-5$, $7=-4$, $8=-3$, $9=-2$, $10=-1$), each again obtained twice by reversing the pairs. So every nonzero residue occurs exactly $\lambda=2$ times among the $20$ ordered differences, and the translates form a symmetric BIBD$(11,11,5,5,2)$; check $\lambda(v-1)=20=k(k-1)$.
7 $\frac{k}{\lambda v}=\frac38$: $\hat\tau=(1.5,-0.75,0.375,-1.125)$. SS $=\frac38(16+4+1+9)=11.25$ on $3$ df. $\operatorname{Var}(\hat\tau_1-\hat\tau_2)=\frac{2\cdot3}{8}\sigma^2=0.75\sigma^2$.
8 $(22,7,2)$: $v$ even and $k-\lambda=5$ not a square, so Theorem 8.34(2) rules it out (it also satisfies $\lambda(v-1)=42=k(k-1)$, so only this test catches it). $(16,6,2)$: $\lambda(v-1)=30=k(k-1)$, $k-\lambda=4=2^2$ — no obstruction (such designs exist). Two blocks share $\lambda=2$ treatments.
9 $\lambda=\frac{r(k-1)}{v-1}$, so $E=\frac{\lambda v}{rk}=\frac{v(k-1)}{k(v-1)}$. Fano: $\frac{7\cdot2}{3\cdot6}=\frac79$; Example 3: $\frac{4\cdot2}{3\cdot3}=\frac89$. As $v\to\infty$, $E\to\frac{k-1}{k}$: with blocks of $2$, at most half the information of complete blocks.
III
$2^k$ factorial experiments.
  • 1 Write the table of signs of $A$, $B$, $AB$ for the four combinations of a $2^2$ and verify that the three columns are mutually orthogonal contrasts.
  • 2 In a $2^2$ experiment with $r=3$, the totals are $(1)=20$, $a=30$, $b=25$, $ab=45$. Estimate $A$, $B$, $AB$ and give their sums of squares.
  • 3 For Exercise III.2, compute the effect of $A$ at each level of $B$ and check your value of $\widehat{AB}$.
  • 4 Apply Yates' algorithm to the totals ($r=1$) $(1)=10$, $a=14$, $b=12$, $ab=20$, $c=11$, $ac=15$, $bc=13$, $abc=25$, and give all effect estimates and sums of squares.
  • 5 Prove that in a $2^k$ factorial the sign column of $X$ times that of $Y$ is the sign column of $X\triangle Y$ (e.g. $AB\cdot BC=AC$).
  • 6 A $2^4$ factorial is run once ($r=1$). How many main effects, two-, three- and four-factor interactions are there? If the three- and four-factor interactions are assumed negligible and pooled as error, how many error df remain?
  • 7 Show that $\operatorname{Var}(\hat X)=\sigma^2/(r2^{k-2})$, and find the number of replicates of a $2^3$ needed for every effect to have standard error at most $\sigma/4$.
  • 8 Explain why a one-factor-at-a-time experiment (vary $A$ with $B,C$ fixed, then vary $B$, then $C$) with the same number of runs as a $2^3$ gives less precise main effects and cannot detect interactions.
1 Order $(1),a,b,ab$: $A=(-,+,-,+)$, $B=(-,-,+,+)$, $AB=(+,-,-,+)$. Each sums to $0$; $A\cdot B=1-1-1+1=0$, $A\cdot AB=-1-1+1+1=0$, $B\cdot AB=-1+1-1+1=0$.
2 Contrasts: $[A]=45+30-25-20=30$, $[B]=45+25-30-20=20$, $[AB]=45-30-25+20=10$. Divide by $r2^{k-1}=6$: $\hat A=5$, $\hat B=3.33$, $\widehat{AB}=1.67$. SS $=$ contrast$^2/12$: $75$, $33.33$, $8.33$.
3 At low $B$: $(30-20)/3=3.33$; at high $B$: $(45-25)/3=6.67$. Half the difference: $1.67=\widehat{AB}$ ✓; the average $5=\hat A$ ✓.
4 Columns: (1) $24,32,26,38,4,8,4,12$; (2) $56,64,12,16,8,12,4,8$; (3) $120,28,20,12,8,4,4,4$. Effects (divide by $4$): $A=7$, $B=5$, $AB=3$, $C=2$, $AC=1$, $BC=1$, $ABC=1$. SS (square, divide by $8$): $98,50,18,8,2,2,2$.
5 $s_X(t)s_Y(t)=\prod_{\ell\in X}(2t_\ell-1)\prod_{\ell\in Y}(2t_\ell-1)$; letters in both $X$ and $Y$ appear squared, and $(2t_\ell-1)^2=1$. What remains is $\prod_{\ell\in X\triangle Y}(2t_\ell-1)=s_{X\triangle Y}(t)$.
6 $\binom41=4$, $\binom42=6$, $\binom43=4$, $\binom44=1$ (total $15$). Pooling the last two classes: $5$ error df for testing the $10$ main effects and two-factor interactions.
7 Theorem 8.36(2). For $k=3$: $\operatorname{Var}=\sigma^2/(2r)\le\sigma^2/16$ iff $r\ge8$.
8 With $8$ runs one-at-a-time, say $2$ runs at each of four settings (base, $A$ high, $B$ high, $C$ high), $\hat A$ is a difference of two means of $2$: variance $\sigma^2$; the $2^3$ gives $\sigma^2/2$ (all $8$ runs are used for every effect). And each factor is varied only at one setting of the others, so $AB$ is not estimable at all: the conclusion about $A$ may hold only at the base levels of $B$ and $C$.
IV
Confounding.
  • 1 Arrange a $2^3$ factorial in two blocks of four confounding $AC$.
  • 2 Arrange a $2^4$ factorial in two blocks of eight confounding $ABCD$, and list the principal block.
  • 3 In a $2^4$ factorial in four blocks, the principal block is $\{(1),bc,abd,acd\}$. Find the confounded effects. Write down the block containing $a$.
  • 4 Arrange a $2^5$ factorial in four blocks of eight confounding $ABC$ and $CDE$. Which other effect is confounded? List the principal block.
  • 5 A $2^3$ factorial is run in three replicates, with $AB$, $AC$ and $BC$ confounded in replicates I, II and III. Give the relative information on each effect and the ANOVA degrees of freedom.
  • 6 Prove that if a $2^5$ factorial is arranged in eight blocks of four, some main effect or two-factor interaction must be confounded.
  • 7 A $2^3$ factorial is run in three replicates of two blocks each, with $ABC$ confounded in every replicate. Give the ANOVA degrees of freedom and the variance of $\hat A$.
1 $L=t_A+t_C$: $\{(1),b,ac,abc\}$ and $\{a,ab,c,bc\}$.
2 Even number of letters: $\{(1),ab,ac,bc,ad,bd,cd,abcd\}$; the other block is the odd ones $\{a,b,c,d,abc,abd,acd,bcd\}$.
3 An effect is confounded iff its parity is $0$ on the whole principal block. $bc$ has even overlap with $ABC$, $BCD$, $AD$; $abd$: $|\{a,b\}|=2$ for $ABC$, $|\{b,d\}|=2$ for $BCD$, $|\{a,d\}|=2$ for $AD$ ✓; $acd$ similarly. So $ABC$, $BCD$ and their generalised interaction $AD$ (as in Example 6(b)). Block of $a$: $a\cdot\{(1),bc,abd,acd\}=\{a,abc,bd,cd\}$.
4 $ABC\cdot CDE=ABDE$. Principal block (even parity for both $ABC$ and $CDE$): $\{(1),ab,de,abde,acd,bcd,ace,bce\}$ (closed under multiplication: e.g. $ab\cdot acd=bcd$).
5 $A,B,C,ABC$: full information. $AB,AC,BC$: each unconfounded in $2$ of $3$ replicates, relative information $\frac23$. df: total $23$, blocks $5$, treatments $7$, error $11$.
6 The $7$ confounded effects are the nonzero vectors of a $3$-dimensional subspace $S$ of $\mathbb F_2^{\,5}$. For each letter $\ell$, the map $s\mapsto s_\ell$ is linear on $S$, so it is either identically $0$ or equal to $1$ on exactly half of $S$, i.e. on $4$ vectors. So the total number of letters in the $7$ effects is at most $5\cdot4=20$, and the average is at most $20/7\lt3$: some confounded effect has at most $2$ letters.
7 Total $23$; blocks $6-1=5$; effects $A,B,C,AB,AC,BC$: $6$; error $23-5-6=12$. $\operatorname{Var}(\hat A)=\frac{\sigma^2}{3\cdot2}=\frac{\sigma^2}{6}$.
🎯 Exam corner CSIR Part BCSIR Part CGATE ST

E1. CSIR Part B GATE ST A balanced incomplete block design has $v=9$ treatments in blocks of size $k=3$, every pair of treatments occurring together in exactly one block. The number of blocks is (A) $9$; (B) $12$; (C) $24$; (D) $36$.

Idea $r=\frac{\lambda(v-1)}{k-1}=\frac82=4$, then $b=\frac{vr}{k}=\frac{36}{3}=12$.
Distractors (A) would be a symmetric design, impossible here since $r\ne k$; (D) is $vr$, forgetting to divide by $k$; (C) has no derivation. (The design exists: the affine plane of order $3$.)
Answer (B).

E2. CSIR Part B A $2^4$ factorial experiment is arranged in four blocks of four plots by confounding $ABC$ and $ACD$. The third effect confounded with blocks is (A) $BD$; (B) $ABD$; (C) $BCD$; (D) $AD$.

Idea The generalised interaction: $ABC\cdot ACD=A^2BC^2D=BD$ (Theorem 8.38(2)).
Check The principal block $\{(1),abd,bcd,ac\}$ has even parity for $BD$ ($bd$: $2$ letters, $abd\to\{b,d\}$, $bcd\to\{b,d\}$, $ac\to\varnothing$) ✓; $ABD$ is odd on $ac$, $BCD$ odd on $ac$, $AD$ odd on $ac$.
Answer (A).

E3. CSIR Part B For the BIBD with parameters $v=b=7$, $r=k=3$, $\lambda=1$, the efficiency factor relative to a randomised block design with the same number of replicates and the same error variance is (A) $\frac79$; (B) $\frac13$; (C) $\frac37$; (D) $\frac67$.

Formula $E=\frac{\lambda v}{rk}=\frac{7}{9}$ (Theorem 8.33(4)): $\operatorname{Var}(\hat\tau_i-\hat\tau_{i'})=\frac{2k\sigma^2}{\lambda v}=\frac67\sigma^2$ against $\frac{2\sigma^2}{3}$ for the RBD.
Distractors (D) is the variance factor $\frac{2k}{\lambda v}$, not the efficiency; (B) is $\lambda/r$; (C) is $k/v$.
Answer (A).

E4. CSIR Part C Let $N$ be the incidence matrix of a BIBD$(v,b,r,k,\lambda)$. Which of the following are necessarily true? (One or more options may be correct.) (A) $b\ge v+r-k$. (B) $r\ge k$. (C) $NN^{\mathsf T}$ is singular. (D) Any two blocks have exactly $\lambda$ treatments in common.

(A) True $b-v=\frac vk(r-k)\ge r-k$, since $r\ge k$ and $v\gt k$ (Exercise II.5).
(B) True Fisher's inequality $b\ge v$ and $bk=vr$.
(C) False $\det NN^{\mathsf T}=rk(r-\lambda)^{v-1}\gt0$ (Theorem 8.32(2)).
(D) False True only for symmetric designs. In the unreduced design with $v=4$, $k=2$, blocks $\{1,2\}$ and $\{3,4\}$ share no treatment, $\{1,2\}$ and $\{1,3\}$ share one.
Answer (A) and (B).

E5. CSIR Part C Consider block designs with $v$ treatments and C-matrix $C=R-NK^{-1}N^{\mathsf T}$. Which of the following are true? (One or more options may be correct.) (A) The design is connected if and only if $\operatorname{rank}C=v-1$. (B) Every complete block design is orthogonal. (C) A randomised block design is orthogonal. (D) In a BIBD all elementary treatment contrasts $\tau_i-\tau_{i'}$ are estimated with the same variance.

(A) True Theorem 8.30: $\operatorname{rank}C=v-(\text{number of components})$.
(B) False Completeness only says $n_{ij}\ge1$. With blocks of sizes $4$ and $3$, treatment 1 twice in block 1, $n_{11}=2\ne r_1k_1/n=12/7$.
(C) True $N=J$ and $rk^{\mathsf T}/n=bv\,J/(bv)=J$.
(D) True $\operatorname{Var}(\hat\tau_i-\hat\tau_{i'})=2k\sigma^2/(\lambda v)$ for every pair (Theorem 8.33(2)).
Answer (A), (C) and (D).

E6. CSIR Part C A $2^3$ factorial experiment is conducted in two replicates; each replicate is divided into two blocks of four plots with $ABC$ confounded in both replicates. Which of the following are true? (One or more options may be correct.) (A) $ABC$ cannot be estimated from the intrablock analysis. (B) The error sum of squares has $6$ degrees of freedom. (C) The block containing $(1)$ is $\{(1),ab,ac,bc\}$. (D) The variance of the estimate of the main effect $A$ is $\sigma^2/8$.

(A) True Its contrast is the difference between block totals in each replicate (total confounding).
(B) True Total $16-1=15$; blocks $4-1=3$; six unconfounded effects; error $15-3-6=6$.
(C) True The combinations with an even number of letters (Example 6(a)).
(D) False $\operatorname{Var}(\hat A)=\frac{\sigma^2}{r2^{k-2}}=\frac{\sigma^2}{2\cdot2}=\frac{\sigma^2}4$: the contrast is a signed sum of $16$ observations, divided by $8$.
Answer (A), (B) and (C).
🧑‍🏫 For the teacher

Suggested plan — 3 lectures × 60 min.

  1. General block designs. Start from the tasting-panel problem. Definition 8.15, Theorem 8.29 (derive the reduced normal equations on the board — it is two lines of block elimination), Theorem 8.30 with the Laplacian picture and Example 1; ask the class to guess the variances in the chain before computing them. Orthogonality (Theorem 8.31) in ten minutes, with the complete-but-not-orthogonal example.
  2. BIBDs. Definition 8.18, Theorem 8.32 by double counting, Fisher's inequality via $\det NN^{\mathsf T}$; Example 2 as a quick-fire class exercise; Theorem 8.33 and Example 3 (have students compute two of the $Q_i$). Project Interactive 1: the Fano plane builds up to equal line thickness, the chain does not, the separate pairs never connect. Symmetric designs (Theorem 8.34) and the difference-set construction with Example 4.
  3. Factorials and confounding. Definition 8.19, Theorem 8.36, Yates' algorithm (Theorem 8.37) on Example 5 using Interactive 2 for the passes; then Definition 8.20 and Theorem 8.38 with the cube — switch between $ABC$, $AB$ and "$AB$ and $AC$" so students see $BC$ disappear too. Example 6(c) as the closing proof, and Example 7 for partial confounding. Finish with E2 and E6.

Discussion questions. (1) Why does Fisher's inequality say that you need at least as many blocks as treatments, even though each block is small? Give an intuitive reason. (2) If you could choose, would you rather lose the $ABC$ interaction completely or lose a quarter of the information on every interaction? What would you need to know about the process? (3) Why is the C-matrix the Laplacian of a graph, and what does "electrical resistance" mean statistically?

Misconceptions to watch for. Using unadjusted treatment means in an incomplete block design; believing integer parameters guarantee a BIBD; claiming any two blocks of every BIBD share $\lambda$ treatments; forgetting generalised interactions; dividing Yates' final column by the wrong constant; confusing "complete", "connected" and "orthogonal".

Suggested homework. I(1, 3, 5, 6), II(2, 3, 6, 7, 8), III(2, 4, 6), IV(3, 4, 5, 6); Exam corner E1, E4, E5.

🔬 Research–Teaching NexusResearch-tutored · optional, about 60 minutes

Research question. Seven coffees are to be ranked by a panel, but no taster can judge more than three cups in a session without fatigue. Design the tasting so that every pair of coffees is compared equally often, and find out by simulation how much precision the incomplete design costs — and what happens if a few tasting sessions are lost.

Why it matters. Sensory panels in the food, wine and cosmetics industries, and consumer studies in many countries, routinely use balanced incomplete block designs because a judge (a block) can only assess a few products. The same designs organise round-robin comparisons in sport, paired comparisons in psychology and the allocation of reviewers to papers. Losing sessions turns a balanced design into an unbalanced one, and the C-matrix tells you exactly how much damage was done.

What to do.

  1. Build the Fano plane from the difference set $\{1,2,4\}$ modulo $7$ (Example 2(d), Interactive 1) and check $\lambda=1$; replicate it twice ($14$ sessions, $r=6$, $\lambda=2$).
  2. Simulate scores $y=\mu+\tau_i+\beta_j+\varepsilon$ with taster effects $\beta_j\sim N(0,1)$, $\varepsilon\sim N(0,0.5^2)$ and chosen coffee effects; fit the intrablock analysis and compare the empirical variance of $\hat\tau_1-\hat\tau_2$ over $2000$ simulations with $2k\sigma^2/(\lambda v)$.
  3. Delete one, two, three sessions at random; compute the new C-matrix and the variances $(e_i-e_{i'})^{\mathsf T}C^-(e_i-e_{i'})$ for all pairs. Which pairs suffer most? Can deleting sessions disconnect the design?
  4. Compare with a design in which each taster scores three coffees chosen at random. How often is that design unbalanced, and how much larger is its average variance?

Tools: Python (numpy) or R; a small function that builds $N$, $C$ and a generalised inverse is all that is needed.

Hand in: a table of pairwise variances for the full and damaged designs, the simulated versus theoretical variance, and a one-paragraph recommendation to the panel organiser.

Going further. Read about A-, D- and E-optimality of block designs (Kiefer's theorem that BIBDs are optimal among designs with the same $v,b,k$) and about the connection between optimal designs and graphs with many spanning trees.

Instructor note: for the doubled Fano plane, $\operatorname{Var}(\hat\tau_i-\hat\tau_{i'})=\frac{2k\sigma^2}{\lambda v}=\frac{6}{14}\sigma^2=\frac37\sigma^2$, i.e. $0.107$ when $\sigma=0.5$; simulations should land within a few per cent of this. Deleting a session lowers the pair counts of three pairs from $2$ to $1$ and raises their variances most; a single lost session cannot disconnect the doubled design, because every pair still meets in the other copy of the Fano plane. Random allocation of coffees to tasters is almost always unbalanced and its worst pair can be much worse than average — a good argument for planned designs.

🧪 Self-check quiz

0/5 correct

Q1. In every BIBD$(v,b,r,k,\lambda)$,

Q2. The C-matrix $C=R-NK^{-1}N^{\mathsf T}$ of any block design

Q3. A $2^3$ factorial is arranged in four blocks of two with $AB$ and $AC$ confounded. Which other effect is necessarily confounded?

Q4. How many passes of Yates' algorithm are needed for a $2^5$ factorial?

Q5. Fisher's inequality for a BIBD states that