8.3 Completely Randomised, Randomised Block and Latin Square Designs
How to compare treatments fairly: randomisation, replication and local control; the completely randomised design and what randomisation alone guarantees; randomised blocks and Latin squares with their models, ANOVA tables and relative efficiencies; and the missing-plot technique for one lost observation.
In 1919 R. A. Fisher arrived at the Rothamsted agricultural station and found seventy years of carefully recorded field trials from which almost nothing could be concluded. A new variety had been planted on the better side of a field, a fertiliser had been tried in a wet year, and no one could say how large a difference had to be before it meant anything. Fisher's diagnosis was that the analysis of an experiment is determined by its design: how the treatments were allotted to the plots decides which comparisons are unbiased, how precise they are, and what the error mean square in an ANOVA table actually estimates. His remedy fits in three words — randomisation, replication, local control — and it is now used in every clinical trial, every industrial experiment and every A/B test on a website.
This section builds the three classical designs on those principles. The completely randomised design (CRD) uses randomisation and replication only; we prove that randomisation by itself makes treatment comparisons unbiased and the error mean square honest, with no normality or independence assumed. The randomised block design (RBD) adds local control in one direction and the Latin square design (LSD) in two; for each we give the model, the ANOVA table and a formula for the relative efficiency, derived from the same randomisation argument. Finally the missing-plot technique rescues an experiment in which one observation has been lost. The linear-model machinery is that of §6.1 ($y=X\beta+\varepsilon$, least squares, estimability, Cochran's theorem) and the analyses are the one- and two-way ANOVAs of §6.2. CSIR NET Part B asks for error degrees of freedom, missing-value estimates and efficiencies; Part C asks which statements about blocking, randomisation and Latin squares are true. Upper points are written $F_{\nu_1,\nu_2,\alpha}$ as in §4.3.
Principles of Experimental Design
An experiment compares $v$ treatments (varieties, doses, machine settings, teaching methods…) by applying them to experimental units — the smallest pieces of material to which a treatment is applied independently (a plot of land, a patient, a batch). A response $y$ is measured on each unit. The design is the rule that decides which unit receives which treatment. Responses of units that received the same treatment still differ: this variation, caused by the inherent differences between units and by measurement noise, is the experimental error, and its variance per unit is the error variance.
We work under unit–treatment additivity: unit $w$ has a fixed unit value $u_w$, and if it receives treatment $i$ its response is $y=u_w+\tau_i+e_w$, where $\tau_i$ is the effect of treatment $i$ and $e_w$ is a technical (measurement) error with mean $0$ and variance $\sigma_e^2$, independent of everything else. Treatments shift responses by the same amount on every unit.
Each principle guards against a different failure. Without randomisation, a new variety planted on the more fertile end of a field "wins" for reasons that have nothing to do with the variety, and no statistical analysis can detect it. Without replication, a single plot per treatment confounds the treatment with the particular plot. Without local control, large differences between units swamp small treatment differences, as Example 3 of §6.2 showed when the fields were ignored. Note what randomisation does not do: it does not reduce variability; it turns an unknown systematic bias into a random error of known average behaviour. Theorem 8.21 makes this precise.
The Completely Randomised Design
In a completely randomised design, $N$ units are available and treatment $i$ is to be applied to $r_i$ of them ($\sum_ir_i=N$). The allocation is chosen at random from all $$\frac{N!}{r_1!\,r_2!\cdots r_v!}$$ possible allocations, for instance by placing the labels in random order (a random permutation of $1,\dots,N$) and giving the first $r_1$ units treatment 1, the next $r_2$ treatment 2, and so on. The model used for the analysis is the one-way classification $$y_{ij}=\mu+\tau_i+\varepsilon_{ij},\qquad i=1,\dots,v,\ j=1,\dots,r_i,$$ with $\varepsilon_{ij}$ iid $N(0,\sigma^2)$, and the analysis is the one-way ANOVA of §6.2: treatments on $v-1$ and error on $N-v$ degrees of freedom, $F=\text{MSTr}/\text{MSE}$.
The normal model is convenient, but where does it come from? The units are not a random sample from anything — they are the particular plots of one field. The next theorem shows that the randomisation itself produces an "error" with exactly the first- and second-moment properties that the ANOVA uses. It is the reason the analysis of a properly randomised experiment can be trusted even when the units are far from iid.
Let $N=vr$ units with unit values $u_1,\dots,u_N$ be allotted by a CRD with $r$ units per treatment, under unit–treatment additivity (Definition 8.10). Write $\bar u$ for the mean of the unit values, $$S_u^2=\frac1{N-1}\sum_{w=1}^N(u_w-\bar u)^2,$$ and let $\bar y_i$ be the mean response of the units given treatment $i$, $\bar\tau=\frac1v\sum\tau_i$. Then, with expectations taken over the random allocation and the technical errors:
- $E[\bar y_i-\bar y_{i'}]=\tau_i-\tau_{i'}$: every treatment difference is estimated unbiasedly;
- $\operatorname{Var}(\bar y_i-\bar y_{i'})=\dfrac{2(S_u^2+\sigma_e^2)}{r}$;
- $E[\text{MSE}]=S_u^2+\sigma_e^2$ and $E[\text{MSTr}]=S_u^2+\sigma_e^2+\dfrac{r\sum_i(\tau_i-\bar\tau)^2}{v-1}$.
So the CRD behaves exactly like the normal one-way model with error variance $\sigma^2=S_u^2+\sigma_e^2$, as far as means and variances are concerned.
Proof. Let $\bar u_i$ and $\bar e_i$ be the averages of the unit values and of the technical errors over the units given treatment $i$, so that $\bar y_i=\bar u_i+\tau_i+\bar e_i$. By symmetry of the randomisation, the set of units receiving treatment $i$ is equally likely to be any $r$-subset of the $N$ units: it is a simple random sample without replacement. By the theorem on the sample mean under SRSWOR (§8.1), $$E[\bar u_i]=\bar u,\qquad\operatorname{Var}(\bar u_i)=\Bigl(1-\frac rN\Bigr)\frac{S_u^2}{r}=\frac{v-1}{v}\cdot\frac{S_u^2}{r}.$$ For the covariance note that $\sum_i\bar u_i=v\bar u$ is a constant, so $$0=\operatorname{Var}\Bigl(\sum_i\bar u_i\Bigr)=v\operatorname{Var}(\bar u_1)+v(v-1)\operatorname{Cov}(\bar u_1,\bar u_2),$$ and therefore $\operatorname{Cov}(\bar u_i,\bar u_{i'})=-\operatorname{Var}(\bar u_1)/(v-1)=-S_u^2/N$ for $i\ne i'$. The $\bar e_i$ are independent of the allocation and of each other, with mean $0$ and variance $\sigma_e^2/r$.
(1) $E[\bar y_i-\bar y_{i'}]=(\bar u-\bar u)+\tau_i-\tau_{i'}+0$. (2) Using the variance and covariance just found, $$\operatorname{Var}(\bar u_i-\bar u_{i'})=2\operatorname{Var}(\bar u_1)-2\operatorname{Cov}(\bar u_1,\bar u_2)$$ $$=\frac{2(v-1)S_u^2}{vr}+\frac{2S_u^2}{vr}=\frac{2S_u^2}{r},$$ and the errors add $2\sigma_e^2/r$.
(3) Since $\bar y_i-\bar y=(\tau_i-\bar\tau)+(\bar u_i-\bar u)+(\bar e_i-\bar e)$ and the last two terms have mean zero, $$E[\text{SSTr}]=r\sum_i(\tau_i-\bar\tau)^2+r\sum_i\operatorname{Var}(\bar u_i)+rE\sum_i(\bar e_i-\bar e)^2=r\sum_i(\tau_i-\bar\tau)^2+(v-1)(S_u^2+\sigma_e^2).$$ For the total sum of squares, $y_w-\bar y=(u_w-\bar u)+(\tau_{t(w)}-\bar\tau)+(e_w-\bar e)$, where $t(w)$ is the treatment of unit $w$. The first term contributes the constant $(N-1)S_u^2$, the second $r\sum(\tau_i-\bar\tau)^2$, the third $(N-1)\sigma_e^2$ in expectation, and the cross term $$2\sum_w(u_w-\bar u)(\tau_{t(w)}-\bar\tau)=2r\sum_i(\tau_i-\bar\tau)(\bar u_i-\bar u)$$ has mean zero (cross terms with the errors also have mean zero). Subtracting $E[\text{SSTr}]$ from $E[\text{SST}]$ gives $E[\text{SSE}]=(N-v)(S_u^2+\sigma_e^2)$. Divide by the degrees of freedom. ∎
Randomisation also gives an exact test that needs no model at all: under $H_0:\tau_1=\cdots=\tau_v$ every allocation would have produced the same responses, so the observed $F$ can be compared with its distribution over all allocations (Fisher's randomisation test, a permutation test as in §5.3). The usual $F_{v-1,N-v}$ distribution is a close approximation to that permutation distribution in all but very small experiments.
Consider a CRD with error variance $\sigma^2$ and $N$ units in all.
- (Equal replication.) The average variance of the $\binom v2$ elementary contrasts $\bar y_i-\bar y_{i'}$ is $\frac{2\sigma^2}{v}\sum_i\frac1{r_i}$; subject to $\sum r_i=N$ it is smallest when all $r_i=N/v$.
- (Square-root rule.) If a control (treatment $0$, $r_0$ units) is to be compared with $m$ test treatments ($r$ units each), $r_0+mr=N$, then $\operatorname{Var}(\bar y_0-\bar y_i)=\sigma^2\left(\frac1{r_0}+\frac1r\right)\ge\frac{\sigma^2(1+\sqrt m)^2}{N}$, with equality iff $r_0=r\sqrt m$.
- (Replication for power.) To detect a difference $\delta$ between two treatments with a two-sided level-$\alpha$ test and power $1-\beta$, about $$r\ge\frac{2\sigma^2(z_{\alpha/2}+z_\beta)^2}{\delta^2}$$ replicates per treatment are needed (normal approximation, $\sigma$ known; $z_\alpha=\Phi^{-1}(1-\alpha)$).
Proof. (1) Each $1/r_i$ occurs in $v-1$ of the $\binom v2$ pairs, so the average of $\sigma^2(1/r_i+1/r_{i'})$ is $\frac{(v-1)\sigma^2\sum1/r_i}{v(v-1)/2}=\frac{2\sigma^2}{v}\sum\frac1{r_i}$. By the Cauchy–Schwarz inequality $\left(\sum1/r_i\right)\left(\sum r_i\right)\ge v^2$, so $\sum1/r_i\ge v^2/N$, with equality iff all $r_i$ are equal. (2) By Cauchy–Schwarz applied to the vectors $(1/\sqrt{r_0},1/\sqrt r)$ and $(\sqrt{r_0},\sqrt{mr})$, $$\left(\frac1{r_0}+\frac1r\right)(r_0+mr)\ge(1+\sqrt m)^2,$$ with equality iff the vectors are proportional, i.e. $\frac{1/\sqrt{r_0}}{\sqrt{r_0}}=\frac{1/\sqrt r}{\sqrt{mr}}$, i.e. $r_0=r\sqrt m$. (3) The test rejects when $|\bar y_1-\bar y_2|\gt z_{\alpha/2}\sqrt{2\sigma^2/r}$. If the true difference is $\delta\gt0$, the power is at least $P\bigl(Z\gt z_{\alpha/2}-\delta/\sqrt{2\sigma^2/r}\bigr)$, which is $\ge1-\beta$ iff $\delta/\sqrt{2\sigma^2/r}\ge z_{\alpha/2}+z_\beta$. ∎
The Randomised Block Design
The units are grouped into $b$ blocks of $v$ units each, chosen so that units within a block are as alike as possible. Every treatment is applied to exactly one unit in every block, the allocation within each block being a random permutation, chosen independently for different blocks. The model is the additive two-way classification $$y_{ij}=\mu+\tau_i+\beta_j+\varepsilon_{ij},\qquad i=1,\dots,v,\ j=1,\dots,b,$$ with $\sum\tau_i=\sum\beta_j=0$ and $\varepsilon_{ij}$ iid $N(0,\sigma^2)$, where $\beta_j$ is the effect of block $j$. (Here $\beta$ denotes block effects, not the whole parameter vector of §6.1.) With treatment totals $T_i$, block totals $B_j$, grand total $G$ and $\text{CF}=G^2/(bv)$:
| Source | df | SS | $E[\text{MS}]$ |
|---|---|---|---|
| Blocks | $b-1$ | $\text{SSB}=\sum_jB_j^2/v-\text{CF}$ | $\sigma^2+\frac{v\sum\beta_j^2}{b-1}$ |
| Treatments | $v-1$ | $\text{SSTr}=\sum_iT_i^2/b-\text{CF}$ | $\sigma^2+\frac{b\sum\tau_i^2}{v-1}$ |
| Error | $(v-1)(b-1)$ | $\text{SSE}=\text{SST}-\text{SSB}-\text{SSTr}$ | $\sigma^2$ |
| Total | $bv-1$ | $\text{SST}=\sum y_{ij}^2-\text{CF}$ |
The table is the two-way ANOVA with one observation per cell of §6.2, with treatments as one factor and blocks as the other; we do not repeat its proof. What is new is the experimental point of view: blocks are a nuisance factor, introduced only to remove variation, and the questions are what the analysis assumes about them and how much they have gained.
- (Estimation.) Every treatment contrast $\sum c_i\tau_i$ ($\sum c_i=0$) is estimable, with BLUE $\sum c_i\bar y_{i\cdot}$ and variance $\sigma^2\sum c_i^2/b$; in particular $\operatorname{Var}(\bar y_{i\cdot}-\bar y_{i'\cdot})=2\sigma^2/b$. These estimates do not involve the block effects at all.
- (Test.) $H_0:\tau_1=\cdots=\tau_v$ is rejected at level $\alpha$ when $\text{MSTr}/\text{MSE}\gt F_{v-1,(v-1)(b-1),\alpha}$.
- (Random blocks.) If the blocks are a sample from a population of blocks, $\beta_j$ iid $N(0,\sigma_b^2)$ independent of the errors, then $E[\text{MSB}]=\sigma^2+v\sigma_b^2$, the other lines are unchanged, and the treatment test is the same. Two observations in the same block have correlation $\sigma_b^2/(\sigma_b^2+\sigma^2)$.
- (Random block × treatment interaction.) If moreover $y_{ij}$ contains a term $c_{ij}$, iid $N(0,\sigma_{\tau\beta}^2)$ and independent of the rest (the unrestricted mixed model of §6.2), then $$E[\text{MSE}]=\sigma^2+\sigma_{\tau\beta}^2,\quad E[\text{MSTr}]=\sigma^2+\sigma_{\tau\beta}^2+\tfrac{b\sum\tau_i^2}{v-1},\quad E[\text{MSB}]=\sigma^2+\sigma_{\tau\beta}^2+v\sigma_b^2,$$ so $\text{MSTr}/\text{MSE}$ and $\text{MSB}/\text{MSE}$ are still exact $F$-tests; but $\sigma^2$ and $\sigma_{\tau\beta}^2$ cannot be separated without replication within cells.
Proof. (1) Every treatment occurs once in every block, so $\bar y_{i\cdot}=\mu+\tau_i+\bar\beta+\bar\varepsilon_{i\cdot}=\mu+\tau_i+\bar\varepsilon_{i\cdot}$ because $\bar\beta=0$; even without the side condition, $\bar\beta$ is common to all treatments and cancels from every contrast. Hence $E[\sum c_i\bar y_{i\cdot}]=\sum c_i\tau_i$, and since the $\bar\varepsilon_{i\cdot}$ are independent with variance $\sigma^2/b$, $\operatorname{Var}=\sigma^2\sum c_i^2/b$. The fitted values of the additive model are $\bar y_{i\cdot}+\bar y_{\cdot j}-\bar y_{\cdot\cdot}$ (§6.2), so $\sum c_i\bar y_{i\cdot}$ is the least squares estimate, which is the BLUE by the Gauss–Markov theorem (§6.1). (2) is the $F$-test of the two-way ANOVA with one observation per cell (§6.2).
(3), (4) We prove (4); (3) is the case $\sigma_{\tau\beta}^2=0$. Now $y_{ij}=\mu+\tau_i+\beta_j+c_{ij}+\varepsilon_{ij}$. The residual array $y_{ij}-\bar y_{i\cdot}-\bar y_{\cdot j}+\bar y_{\cdot\cdot}$ contains only the double-centred $c$ and $\varepsilon$ terms; for iid terms the double-centred array has expected squared length $(v-1)(b-1)$ times the variance (it is the residual of the additive model fitted to pure noise, §6.1), so $E[\text{SSE}]=(v-1)(b-1)(\sigma^2+\sigma_{\tau\beta}^2)$. Next $\bar y_{i\cdot}-\bar y_{\cdot\cdot}=\tau_i+(\bar c_{i\cdot}-\bar c_{\cdot\cdot})+(\bar\varepsilon_{i\cdot}-\bar\varepsilon_{\cdot\cdot})$ — the $\beta_j$ cancel because every treatment mean contains the same $\bar\beta$ — and $\bar c_{i\cdot}+\bar\varepsilon_{i\cdot}$ are iid with variance $(\sigma_{\tau\beta}^2+\sigma^2)/b$, so $E[\text{SSTr}]=b\sum\tau_i^2+(v-1)(\sigma^2+\sigma_{\tau\beta}^2)$. Similarly $\bar y_{\cdot j}-\bar y_{\cdot\cdot}=(\beta_j-\bar\beta)+(\bar c_{\cdot j}-\bar c_{\cdot\cdot})+(\bar\varepsilon_{\cdot j}-\bar\varepsilon_{\cdot\cdot})$ with iid summands of variance $\sigma_b^2+(\sigma_{\tau\beta}^2+\sigma^2)/v$, giving $E[\text{SSB}]=(b-1)(\sigma^2+\sigma_{\tau\beta}^2+v\sigma_b^2)$. In this balanced layout the three sums of squares come from mutually orthogonal projections of a normal vector whose covariance is constant on each of the three subspaces, so each divided by its expected mean square times its df is $\chi^2$, independently (the argument used for the expected mean squares of the balanced two-way layout in §6.2); under each null hypothesis the ratio is therefore exactly $F$. ∎
Part (4) agrees with the mixed-model rule of §6.2 — the fixed factor is tested against the interaction mean square — because with one observation per cell "error" and "interaction" are the same line of the table. If instead the blocks are fixed and interact with treatments, MSE overestimates $\sigma^2$ and the treatment test becomes conservative; the treatment means then estimate effects averaged over these particular blocks.
Let $b$ blocks of $v$ units be allotted by an RBD under unit–treatment additivity, with unit values $u_{jw}$ (block $j$, unit $w$). Let $\bar u_j$ be the mean unit value of block $j$, $\bar u$ the overall mean, $$S_j^2=\frac1{v-1}\sum_w(u_{jw}-\bar u_j)^2,\qquad S_w^2=\frac1b\sum_jS_j^2,$$ and $S_u^2$ the variance (divisor $bv-1$) of all $bv$ unit values. Then over the randomisation:
- $E[\bar y_{i\cdot}-\bar y_{i'\cdot}]=\tau_i-\tau_{i'}$ and $\operatorname{Var}(\bar y_{i\cdot}-\bar y_{i'\cdot})=2(S_w^2+\sigma_e^2)/b$;
- $E[\text{MSE}]=S_w^2+\sigma_e^2$ and $E[\text{MSTr}]=S_w^2+\sigma_e^2+b\sum_i(\tau_i-\bar\tau)^2/(v-1)$;
- $\displaystyle\hat\sigma^2_{\text{CRD}}=\frac{(b-1)\text{MSB}+b(v-1)\text{MSE}}{bv-1}$ is an unbiased estimator of $S_u^2+\sigma_e^2$, the error variance that a CRD with $b$ replicates on the same units would have had (Theorem 8.21). The relative efficiency of the RBD with respect to the CRD, $(S_u^2+\sigma_e^2)/(S_w^2+\sigma_e^2)$, is therefore estimated by $$\widehat{\text{RE}}=\frac{(b-1)\text{MSB}+b(v-1)\text{MSE}}{(bv-1)\text{MSE}};$$
- $\widehat{\text{RE}}-1=\dfrac{(b-1)(\text{MSB}-\text{MSE})}{(bv-1)\text{MSE}}$, so blocking is estimated to have helped if and only if $\text{MSB}\gt\text{MSE}$.
Proof. Take $\sigma_e^2=0$ first. Write $d_{ji}=u_{j,w}-\bar u_j$ for the unit $w$ of block $j$ that receives treatment $i$. Within block $j$, $(d_{j1},\dots,d_{jv})$ is a random permutation of the $v$ deviations of that block, which have mean $0$ and population variance $\frac{v-1}{v}S_j^2$; two distinct positions of a random permutation are two draws without replacement, with covariance $-\frac{1}{v-1}\cdot\frac{v-1}{v}S_j^2=-S_j^2/v$. Blocks are randomised independently.
(1) $\bar y_{i\cdot}=\bar u+\tau_i+D_i$ with $D_i=\frac1b\sum_jd_{ji}$, $E[D_i]=0$, and $$\operatorname{Var}(D_i-D_{i'})=\frac1{b^2}\sum_j\Bigl[2\cdot\tfrac{v-1}{v}S_j^2+2\cdot\tfrac{S_j^2}{v}\Bigr]=\frac{2}{b^2}\sum_jS_j^2=\frac{2S_w^2}{b}.$$
(2) $\bar y_{i\cdot}-\bar y_{\cdot\cdot}=(\tau_i-\bar\tau)+D_i$ (note $\sum_iD_i=0$). Since $\operatorname{Var}(D_i)=\frac1{b^2}\sum_j\frac{v-1}{v}S_j^2$, $$E[\text{SSTr}]=b\sum_i(\tau_i-\bar\tau)^2+b\sum_i\operatorname{Var}(D_i)$$ $$=b\sum_i(\tau_i-\bar\tau)^2+(v-1)S_w^2.$$ The within-block sum of squares is $$\sum_{ij}(y_{ij}-\bar y_{\cdot j})^2=\sum_{ij}(\tau_i-\bar\tau+d_{ji})^2=b\sum_i(\tau_i-\bar\tau)^2+2\sum_i(\tau_i-\bar\tau)\,bD_i+b(v-1)S_w^2,$$ whose expectation is $b\sum(\tau_i-\bar\tau)^2+b(v-1)S_w^2$. Since $\text{SSE}$ is the within-block sum of squares minus SSTr, $E[\text{SSE}]=(b-1)(v-1)S_w^2$.
(3) Each block contains every treatment once, so $\bar y_{\cdot j}=\bar u_j+\bar\tau$ and $\text{SSB}=v\sum_j(\bar u_j-\bar u)^2$ is a constant. The between/within decomposition of the unit values reads $$(bv-1)S_u^2=v\sum_j(\bar u_j-\bar u)^2+b(v-1)S_w^2=\text{SSB}+b(v-1)S_w^2,$$ and with $E[\text{MSE}]=S_w^2$ this gives $E[(b-1)\text{MSB}+b(v-1)\text{MSE}]=(bv-1)S_u^2$. With technical errors, $E[\text{SSB}]$ gains $(b-1)\sigma_e^2$ and $E[\text{MSE}]$ gains $\sigma_e^2$, so the left side gains $(bv-1)\sigma_e^2$, as required; the other statements gain $\sigma_e^2$ in the same way as in Theorem 8.21. By Theorem 8.21(2), a CRD with $r=b$ would give $\operatorname{Var}=2(S_u^2+\sigma_e^2)/b$, so the ratio of variances is $(S_u^2+\sigma_e^2)/(S_w^2+\sigma_e^2)$.
(4) The numerator of $\widehat{\text{RE}}-1$ is $(b-1)\text{MSB}+b(v-1)\text{MSE}-(bv-1)\text{MSE}$, and $b(v-1)-(bv-1)=1-b$. ∎
The estimator in (3) is not the naive "pool blocks with error", $(\text{SSB}+\text{SSE})/(v(b-1))$: in a CRD on the same units the treatment line would also have absorbed unit variation, so the treatment sum of squares is replaced by its null expectation $(v-1)\text{MSE}$ and the total is divided by all $bv-1$ degrees of freedom. Blocking also costs error degrees of freedom — $(v-1)(b-1)$ instead of $v(b-1)$ — and Fisher suggested multiplying $\widehat{\text{RE}}$ by $\frac{(n_1+1)(n_2+3)}{(n_2+1)(n_1+3)}$, where $n_1,n_2$ are the error df of the RBD and the CRD, to account for the less precise estimate of $\sigma^2$. The factor is close to $1$ unless the experiment is tiny.
| Block 1 | Block 2 | Block 3 | Block 4 | Block 5 | $T_i$ | |
|---|---|---|---|---|---|---|
| V1 | 31 | 40 | 28 | 36 | 41 | 176 |
| V2 | 37 | 41 | 33 | 37 | 48 | 196 |
| V3 | 30 | 37 | 26 | 32 | 36 | 161 |
| V4 | 38 | 43 | 34 | 39 | 49 | 203 |
| $B_j$ | 136 | 161 | 121 | 144 | 174 | $G=736$ |
The Latin Square Design
A field often has two gradients (slope and drainage), a laboratory has two nuisance factors (operator and day), a road test has drivers and cars. A Latin square controls both at once with surprisingly few units.
A Latin square of order $v$ is a $v\times v$ array of $v$ symbols in which each symbol occurs exactly once in each row and exactly once in each column. It is standard (reduced) if its first row and first column are in natural order. In a Latin square design the $v^2$ units are classified by two nuisance factors with $v$ levels each (rows and columns), and the $v$ treatments are allotted according to a Latin square chosen at random — in practice, by taking a standard square and permuting its rows, its columns and the treatment labels at random. With $d(i,j)$ the treatment in row $i$ and column $j$, the model is $$y_{ij}=\mu+\rho_i+\gamma_j+\tau_{d(i,j)}+\varepsilon_{ij},\qquad i,j=1,\dots,v,$$ with $\sum\rho_i=\sum\gamma_j=\sum\tau_k=0$ and $\varepsilon_{ij}$ iid $N(0,\sigma^2)$ (rows, columns and treatments act additively — no interactions).
Two Latin squares of the same order are orthogonal if, when superimposed, every ordered pair of symbols occurs exactly once. There are at most $v-1$ mutually orthogonal Latin squares of order $v$, and $v-1$ exist when $v$ is a prime power; there is no pair of orthogonal Latin squares of order $2$ or $6$ (Euler's problem of the 36 officers, settled by Tarry in 1900), and there is one for every other order (Bose, Shrikhande and Parker, 1959). A superimposed pair gives a Graeco-Latin square, which controls three nuisance factors.
With row totals $R_i$, column totals $C_j$, treatment totals $T_k$, $G=\sum y_{ij}$ and $\text{CF}=G^2/v^2$, let $$\text{SSR}=\frac{\sum R_i^2}{v}-\text{CF},\quad\text{SSC}=\frac{\sum C_j^2}{v}-\text{CF},\quad\text{SSTr}=\frac{\sum T_k^2}{v}-\text{CF},$$ and $\text{SSE}=\text{SST}-\text{SSR}-\text{SSC}-\text{SSTr}$. Then:
- $\text{SST}=\text{SSR}+\text{SSC}+\text{SSTr}+\text{SSE}$ is a sum of squared lengths of mutually orthogonal vectors, with degrees of freedom $$v^2-1=(v-1)+(v-1)+(v-1)+(v-1)(v-2);$$
- $E[\text{MSE}]=\sigma^2$, $E[\text{MSTr}]=\sigma^2+\frac{v\sum\tau_k^2}{v-1}$, and similarly for rows and columns;
- under normality the four sums of squares are independent, and $H_0:\tau_1=\cdots=\tau_v$ is rejected when $\text{MSTr}/\text{MSE}\gt F_{v-1,(v-1)(v-2),\alpha}$;
- $\bar y_{(k)}-\bar y_{(k')}$ (treatment means) is the BLUE of $\tau_k-\tau_{k'}$, with variance $2\sigma^2/v$, free of row and column effects.
Proof. Write $\bar y_{i\cdot}$, $\bar y_{\cdot j}$, $\bar y_{(k)}$ for row, column and treatment means and $\bar y$ for the grand mean, and decompose $$y_{ij}-\bar y=(\bar y_{i\cdot}-\bar y)+(\bar y_{\cdot j}-\bar y)+(\bar y_{(d(i,j))}-\bar y)+e_{ij},$$ where $e_{ij}=y_{ij}-\bar y_{i\cdot}-\bar y_{\cdot j}-\bar y_{(d(i,j))}+2\bar y$. Regard the four terms as vectors in $\mathbb R^{v^2}$ (one coordinate per cell). Orthogonality. The row and treatment arrays are orthogonal: as $j$ runs through a row $i$, $d(i,j)$ runs through every treatment exactly once, so $$\sum_{ij}(\bar y_{i\cdot}-\bar y)(\bar y_{(d(i,j))}-\bar y)=\sum_i(\bar y_{i\cdot}-\bar y)\sum_k(\bar y_{(k)}-\bar y)=0.$$ The same argument (each treatment meets each column once, each row meets each column once) gives the other two pairs. The residual array sums to zero along every row: for fixed $i$, $\sum_je_{ij}=v\bar y_{i\cdot}-v\bar y_{i\cdot}-v\bar y-v\bar y+2v\bar y=0$, because $\sum_j\bar y_{\cdot j}=v\bar y$ and $\sum_j\bar y_{(d(i,j))}=\sum_k\bar y_{(k)}=v\bar y$. It also sums to zero along every column and over the $v$ cells of every treatment, by the same computation. Hence $e$ is orthogonal to the three other arrays, which are constant on rows, on columns and on treatments respectively. Squaring and summing gives (1).
Degrees of freedom. The row arrays form a subspace of dimension $v-1$ (constant on rows, total zero), and similarly for columns and treatments; together with the constants these mutually orthogonal subspaces span a space of dimension $1+3(v-1)=3v-2$, and the residual lies in its orthogonal complement, of dimension $v^2-3v+2=(v-1)(v-2)$. Expectations. Substituting the model, every row and every column contains each treatment once, so $\bar y_{(k)}=\mu+\tau_k+\bar\varepsilon_{(k)}$: the $\rho_i$ and $\gamma_j$ average to zero. Hence $\bar y_{(k)}-\bar y=\tau_k+(\bar\varepsilon_{(k)}-\bar\varepsilon)$ and $E[\text{SSTr}]=v\sum\tau_k^2+(v-1)\sigma^2$; the residual array contains only $\varepsilon$'s (all parameters cancel), and is the residual of a linear model of rank $3v-2$, so $E[\text{SSE}]=(v-1)(v-2)\sigma^2$ (unbiased estimation of $\sigma^2$, §6.1). This gives (2) and (4) (the variance of a difference of two treatment means, each an average of $v$ independent errors, is $2\sigma^2/v$; it is the least squares estimate because the fitted values are $\bar y_{i\cdot}+\bar y_{\cdot j}+\bar y_{(k)}-2\bar y$). (3) follows from Cochran's theorem (§6.1), since the projections onto the four orthogonal subspaces have ranks adding up to $v^2-1$. ∎
The price of controlling two factors is error degrees of freedom: $(v-1)(v-2)$ is $0$ for $v=2$ (no test is possible), $2$ for $v=3$, $6$ for $v=4$ and $12$ for $v=5$. Small squares are therefore often replicated, or used only when the row and column effects are known to be large. A Graeco-Latin square of order $v$ leaves $(v-1)(v-3)$ error df. Randomisation matters here too: there are $12$ Latin squares of order $3$ (all obtainable from one by permuting rows, columns and symbols) and $576$ of order $4$, of which $4$ are standard; permuting the rows, columns and labels of the cyclic $4\times4$ square reaches only $432$ of the $576$, which is why the recipe starts from a standard square chosen at random.
From the ANOVA of a Latin square, the error variances that other designs on the same $v^2$ units would have had are estimated by $$\hat\sigma^2_{\text{CRD}}=\frac{\text{MSR}+\text{MSC}+(v-1)\text{MSE}}{v+1},\qquad\hat\sigma^2_{\text{RBD, rows as blocks}}=\frac{\text{MSC}+(v-1)\text{MSE}}{v},$$ and $\hat\sigma^2_{\text{RBD, columns as blocks}}=\frac{\text{MSR}+(v-1)\text{MSE}}{v}$. The relative efficiencies of the LSD are these estimates divided by MSE.
Proof (sketch). The argument of Theorem 8.24 applies with two nuisance classifications. Every row and every column contains each treatment once, so the row and column sums of squares are fixed by the units (plus technical error), while the error mean square estimates the variance of the unit values left after removing row and column means. Replace the treatment sum of squares by its null expectation $(v-1)\text{MSE}$; the table then describes the unit values alone, as in a uniformity trial. A CRD would have an error variance equal to the overall variance of the units, estimated by the whole table over all $v^2-1$ df: rows, columns, the treatment line and error contribute $$(v-1)\bigl[\text{MSR}+\text{MSC}+\text{MSE}+(v-2)\text{MSE}\bigr],$$ and dividing by $v^2-1=(v-1)(v+1)$ gives the first formula. An RBD with the rows as blocks would have an error variance equal to the within-row variance of the units. The within-row part of the table (columns, treatments, error) contributes $$(v-1)\bigl[\text{MSC}+\text{MSE}+(v-2)\text{MSE}\bigr]$$ on $v(v-1)$ df, which gives the second formula. The third is symmetric. ∎
| M1 | M2 | M3 | M4 | $R_i$ | |
|---|---|---|---|---|---|
| Op 1 | C 16 | D 17 | A 8 | B 13 | 54 |
| Op 2 | B 19 | C 21 | D 22 | A 16 | 78 |
| Op 3 | A 11 | B 17 | C 16 | D 18 | 62 |
| Op 4 | D 20 | A 15 | B 16 | C 19 | 70 |
| $C_j$ | 66 | 70 | 62 | 66 | $G=264$ |
Missing Plots
A plot is flooded, a patient drops out, a sample is spilled. The design is then unbalanced, its neat orthogonality is gone, and the simple formulae no longer apply. The general remedy is to refit the linear model to the available data (§6.1). Yates (1933) showed that, for one missing value, the same answer is obtained by a trick that keeps the hand formulae: insert the value that makes the error sum of squares as small as possible, analyse the completed table, and make two small corrections.
In a linear model $y=X\beta+\varepsilon$ suppose one observation, in cell $m$ with design row $x_m^{\mathsf T}$, is missing. For a trial value $x$ let $\text{SSE}(x)$ be the residual sum of squares of the completed data (the available data together with $x$ in cell $m$). Then:
- $\text{SSE}(x)\ge\text{RSS}_{\text{avail}}$, the residual sum of squares of the least squares fit to the available data, with equality exactly at $\hat x=x_m^{\mathsf T}\hat\beta$, the fitted value of the missing cell from that fit;
- the least squares estimates from the completed data with $x=\hat x$ coincide with those from the available data, and the error degrees of freedom are one fewer than for complete data;
- for any submodel $H$ (a hypothesis), the residual sum of squares $\text{SSE}_0(\hat x)$ of the completed data under $H$ is at least the correct $\text{RSS}_{0,\text{avail}}$, and it exceeds it by $\text{SSE}_0(\hat x)-\text{SSE}_0(\hat x_0)$, where $\hat x_0$ is the missing-plot value under $H$. Hence the hypothesis sum of squares computed from the completed table is biased upwards by that amount.
Proof. Let $Q(\beta,x)=\sum_{w\ne m}(y_w-x_w^{\mathsf T}\beta)^2+(x-x_m^{\mathsf T}\beta)^2$ be the residual sum of squares of the completed data at the parameter $\beta$. For fixed $\beta$ the last term is minimised (made zero) by $x=x_m^{\mathsf T}\beta$, so $\min_xQ(\beta,x)=R(\beta)$, the available-data residual sum of squares at $\beta$. Therefore $$\min_{\beta,x}Q=\min_\beta R(\beta)=\text{RSS}_{\text{avail}},$$ attained at $\beta=\hat\beta$ and $x=x_m^{\mathsf T}\hat\beta$. Now $\text{SSE}(x)=\min_\beta Q(\beta,x)\ge\min_{\beta,x}Q=\text{RSS}_{\text{avail}}$, with equality when $x=\hat x$ (take $\beta=\hat\beta$). If $\text{SSE}(x)=\text{RSS}_{\text{avail}}$ for some $x$, a minimising $\beta$ gives $R(\beta)\le Q(\beta,x)=\text{RSS}_{\text{avail}}$, so $\beta$ is a least squares solution for the available data; when the fitted value of cell $m$ is the same for all such solutions (i.e. $x_m^{\mathsf T}\beta$ is estimable from the available data, which holds whenever the missing value can be estimated at all), $Q(\beta,x)=\text{RSS}_{\text{avail}}$ forces $x=x_m^{\mathsf T}\beta=\hat x$. This proves (1); and since $Q(\hat\beta,\hat x)$ is the minimum of $Q(\cdot,\hat x)$, $\hat\beta$ is a least squares solution of the completed problem, which is (2). The completed table has one fictitious observation that carries no information, so one error degree of freedom must be removed. (3) Apply (1) to the submodel: $\text{SSE}_0(x)\ge\text{RSS}_{0,\text{avail}}=\text{SSE}_0(\hat x_0)$ for every $x$, in particular for $x=\hat x$. The correct hypothesis sum of squares is $\text{RSS}_{0,\text{avail}}-\text{RSS}_{\text{avail}}=\text{SSE}_0(\hat x_0)-\text{SSE}(\hat x)$, while the completed table gives $\text{SSE}_0(\hat x)-\text{SSE}(\hat x)$. ∎
- (RBD, Yates.) If the observation of treatment $i$ in block $j$ is missing and $T$, $B$, $G$ are the totals of the available observations of that treatment, that block and the whole experiment, then $$\hat x=\frac{vT+bB-G}{(v-1)(b-1)}.$$ The error df become $(v-1)(b-1)-1$, the treatment sum of squares of the completed table must be reduced by $$\text{bias}=\frac{\bigl[B-(v-1)\hat x\bigr]^2}{v(v-1)},$$ and $\operatorname{Var}(\hat\tau_i-\hat\tau_{i'})=\sigma^2\left[\frac2b+\frac{v}{b(b-1)(v-1)}\right]$ for the treatment with the missing value against any other.
- (LSD.) If the observation in row $i$, column $j$ is missing and $R$, $C$, $T$, $G$ are the available totals of that row, column, treatment and the whole square, then $$\hat x=\frac{v(R+C+T)-2G}{(v-1)(v-2)},$$ the error df become $(v-1)(v-2)-1$, and the treatment sum of squares must be reduced by $\dfrac{\bigl[G-R-C-(v-1)T\bigr]^2}{(v-1)^2(v-2)^2}$.
Proof. (RBD) By Theorem 8.27(1), $\hat x$ equals its own fitted value in the completed table, which for the additive model is $\bar y_{i\cdot}+\bar y_{\cdot j}-\bar y_{\cdot\cdot}$ computed with $\hat x$ included: $$\hat x=\frac{T+\hat x}{b}+\frac{B+\hat x}{v}-\frac{G+\hat x}{bv}.$$ Multiplying by $bv$: $\hat x\,(bv-v-b+1)=vT+bB-G$, and $bv-v-b+1=(v-1)(b-1)$. For the bias use Theorem 8.27(3) with $H$: no treatment effects. Under $H$ the model is "blocks only", whose missing-plot value is the mean of the other $v-1$ plots of block $j$, $\hat x_0=B/(v-1)$, and whose residual sum of squares is the within-block sum of squares. As a function of the entry $x$ of block $j$, the within-block sum of squares of that block is $\sum(y-\bar y_{\cdot j})^2$, a quadratic in $x$ with leading coefficient $1-\frac1v$ (the entry $x$ enters with coefficient $1-\frac1v$ in its own deviation and $-\frac1v$ in the $v-1$ others), minimised at $\hat x_0$. So $$\text{SSE}_0(\hat x)-\text{SSE}_0(\hat x_0)=\frac{v-1}{v}\Bigl(\hat x-\frac B{v-1}\Bigr)^2=\frac{\bigl[(v-1)\hat x-B\bigr]^2}{v(v-1)}.$$ The variance formula follows by inverting the normal equations for the incomplete layout (Exercise IV.6 asks for a check of a special case). (LSD) The fitted value is $\bar y_{i\cdot}+\bar y_{\cdot j}+\bar y_{(k)}-2\bar y$, so $$\hat x=\frac{R+\hat x}{v}+\frac{C+\hat x}{v}+\frac{T+\hat x}{v}-\frac{2(G+\hat x)}{v^2};$$ multiplying by $v^2$ gives $\hat x(v^2-3v+2)=v(R+C+T)-2G$. The bias is obtained from Theorem 8.27(3) with the submodel "rows and columns only", by the same kind of computation. ∎
The bias is zero when the missing-plot value under $H$ happens to equal $\hat x$; otherwise ignoring it inflates $F$. With several missing values one solves the equations $\partial\,\text{SSE}/\partial x_l=0$ simultaneously (or iterates the one-value formula, each time treating the other estimates as data), and removes one error df for each — but with modern software one simply fits the model to the available data.
- Identify the design from how the randomisation was restricted: none (CRD), within blocks (RBD), within rows and columns (LSD). The analysis must remove exactly the classifications the randomisation respected.
- Degrees of freedom first. CRD: $v-1$, $N-v$. RBD: $b-1$, $v-1$, $(v-1)(b-1)$. LSD: $3\times(v-1)$ and $(v-1)(v-2)$. Subtract one error df per missing value.
- Totals and the correction factor $\text{CF}=G^2/N$; each classification's SS is $\sum(\text{total})^2/(\text{number of plots in it})-\text{CF}$; error by subtraction.
- Missing value? Estimate it by the formula (RBD: $\frac{vT+bB-G}{(v-1)(b-1)}$; LSD: $\frac{v(R+C+T)-2G}{(v-1)(v-2)}$), complete the table, reduce the error df, and subtract the bias from the treatment SS.
- Efficiency: replace the treatment line by $(v-1)\text{MSE}$ and pool what the other design would not have removed — RBD vs CRD: $\frac{(b-1)M_b+b(v-1)M_e}{(bv-1)M_e}$.
- Randomisation does not reduce variance. It removes bias and makes the error mean square honest (Theorem 8.21); precision comes from replication and local control.
- Blocking can hurt. $\widehat{\text{RE}}\lt1$ whenever $\text{MSB}\lt\text{MSE}$ (e.g. $53/57$ in Example 4), and blocks laid along a gradient make the RBD genuinely worse than a CRD (Interactive 1, blocks = columns). Blocking always costs $b-1$ error df.
- The efficiency formula is not "pool blocks with error". RBD vs CRD is $\frac{(b-1)M_b+b(v-1)M_e}{(bv-1)M_e}$, not $\frac{\text{SSB}+\text{SSE}}{v(b-1)M_e}$: the treatment line is replaced by $(v-1)M_e$ and all $bv-1$ df are used.
- A $2\times2$ Latin square has no error df, a $3\times3$ only $2$; $(v-1)(v-2)$ grows slowly.
- No orthogonal pair of Latin squares of order $6$ (or $2$), although there are pairs for all other orders; for prime powers $v$ there are $v-1$ mutually orthogonal squares.
- An LSD needs as many treatments as rows and columns and assumes no interactions; a row × treatment interaction inflates MSE and can bias other comparisons.
- Each treatment once per block is not enough. Without independent randomisation within blocks (e.g. always in the order A, B, C, D along the row) a within-block trend biases the comparison, and Theorem 8.24 fails.
- Missing plots: two corrections are mandatory. Remove one error df per missing value, and subtract the upward bias $[B-(v-1)\hat x]^2/(v(v-1))$ from the treatment SS; the estimated value is not a datum.
- After a missing value the simple means are not the estimates. $\bar y_{i\cdot}$ of the available plots is biased by the block effects; use the completed means (or least squares), and the larger variance $\sigma^2[\frac2b+\frac{v}{b(b-1)(v-1)}]$.
- Pseudo-replication. Measuring one plot four times gives four numbers but one experimental unit: subsamples estimate measurement error, not the experimental error against which treatments must be compared.
- Random blocks change only the blocks line. With an additive model (or the unrestricted random interaction), the treatment test is $\text{MSTr}/\text{MSE}$ whether blocks are fixed or random.
Exercises
- 1 For each principle — randomisation, replication, local control — describe a realistic experiment that violates it and say what goes wrong.
- 2 A CRD has $5$ treatments with $6,6,5,5,4$ replicates. Write down the sources and degrees of freedom of its ANOVA table.
- 3 How many different allocations are possible in a CRD with $v=3$ treatments and $r=2$ units each? With $v=4$, $r=5$?
- 4 Analyse the CRD with responses: treatment 1: $8,10,9$; treatment 2: $12,14,13,13$; treatment 3: $9,11,10$.
- 5 Prove that, for a fixed total $N$, equal replication minimises the largest variance of an elementary contrast $\bar y_i-\bar y_{i'}$ (not only the average).
- 6 One control and $9$ test treatments are to be compared on $120$ units. Find the square-root allocation and its efficiency relative to equal replication for the control comparisons.
- 7 With $\sigma=2.5$, how many replicates per treatment are needed to detect a difference of $3$ units with a two-sided $5\%$ test and power $80\%$ (normal approximation)?
- 8 Four units have unit values $1,2,3,6$, and two treatments with no effect ($\tau_A=\tau_B$) are allotted by a CRD with $r=2$. List all allocations, the distribution of $\bar y_A-\bar y_B$, and verify Theorem 8.21(2).
- 1 Give the ANOVA sources and degrees of freedom of an RBD with $6$ treatments in $4$ blocks.
- 2 Analyse the RBD (rows treatments, columns blocks): T1: $10,12,9,13$; T2: $14,15,13,16$; T3: $11,14,10,13$. Estimate the efficiency relative to a CRD.
- 3 Show directly from the model that $E[\bar y_{1\cdot}-\bar y_{2\cdot}]=\tau_1-\tau_2$ in an RBD, but that $E[\bar y_{\cdot1}-\bar y_{\cdot2}]$ also involves no treatment effects. Which property of the design is responsible?
- 4 An RBD with $v=5$ treatments and $b=4$ blocks has $\text{MSB}=40$ and $\text{MSE}=8$. Estimate its efficiency relative to a CRD.
- 5 In an RBD with random blocks, give an unbiased estimator of $\sigma_b^2$ and of the correlation between two plots of the same block. Evaluate them for Exercise II.2, treating the blocks as random.
- 6 Explain why, with one observation per cell, the block × treatment interaction cannot be tested in an RBD, and why the treatment test is still valid when the blocks are random (unrestricted model).
- 7 In an RBD with $v=4$ treatments and $b=6$ blocks one value is missing; the available totals are $T=120$ (its treatment), $B=60$ (its block) and $G=545$. Estimate it.
- 8 Two values are missing in an RBD, in different treatments and different blocks: $x$ (treatment $i$, block $j$) and $z$ (treatment $i'$, block $j'$). Show that minimising the error sum of squares leads to $(v-1)(b-1)x+z=vT_i+bB_j-G$ and $x+(v-1)(b-1)z=vT_{i'}+bB_{j'}-G$, where all totals are of available values.
- 1 Write the ANOVA sources and df of a $6\times6$ Latin square design.
- 2 Show that a Latin square design of order $v$ has fewer error df than an RBD with $v$ treatments in $v$ blocks, and find the difference.
- 3 Write down the four standard Latin squares of order $4$.
- 4 Write a cyclic $5\times5$ Latin square and describe a randomisation of it for an experiment with $5$ drivers (rows), $5$ cars (columns) and $5$ fuel additives.
- 5 In a $5\times5$ LSD, $\text{MSR}=30$, $\text{MSC}=12$ and $\text{MSE}=6$. Estimate the efficiency of the LSD relative to a CRD, to an RBD with rows as blocks and to an RBD with columns as blocks.
- 6 In Example 5 the observation of operator 2 on machine 4 (method A, $16$) is lost. Estimate it, and give the corrected treatment and error sums of squares.
- 7 Construct a $3\times3$ Graeco-Latin square, and show that a Graeco-Latin square design of order $v$ leaves $(v-1)(v-3)$ error df.
- 8 Prove that two Latin squares of order $v$ cannot be orthogonal if one is obtained from the other by permuting the symbols only.
- 1 In a CRD with $N=vr$, prove $\operatorname{Cov}(\bar u_i,\bar u_{i'})=-S_u^2/N$ by computing the covariance of two draws from the same random permutation directly (without the constant-sum trick).
- 2 Show that $\widehat{\text{RE}}$ of an RBD with respect to a CRD equals $1+\frac{(b-1)(F_B-1)}{bv-1}$, where $F_B=\text{MSB}/\text{MSE}$, and compute it for Example 3.
- 3 An experimenter has $4$ treatments and $3$ plots per treatment in a CRD, and measures each plot $4$ times. A colleague reports an error mean square with $44$ df. Explain the mistake and give the correct error df for testing treatments.
- 4 (Modelling.) A field of $25$ plots has a fertility gradient from east to west and a drainage gradient from north to south. Five fertilisers are to be compared. Recommend a design, give its ANOVA df, and describe the randomisation.
- 5 In an RBD with a missing value, show that $\text{SSE}(x)=\text{SSE}(\hat x)+\frac{(v-1)(b-1)}{vb}(x-\hat x)^2$.
- 6 For an RBD with $v=2$ treatments, $b$ blocks and the value of treatment 1 in block 1 missing, show that the least squares estimate of $\tau_1-\tau_2$ is the mean of the $b-1$ complete within-block differences, and verify the variance formula of Theorem 8.28.
E1. CSIR Part B GATE ST Six treatments are compared in a $6\times6$ Latin square design. The number of degrees of freedom for error is (A) $25$; (B) $20$; (C) $30$; (D) $24$.
E2. CSIR Part B In a randomised block design with $5$ treatments and $4$ blocks, one observation is missing. The totals of the available observations of its treatment, of its block and of the whole experiment are $46$, $68$ and $330$. The missing-plot estimate is (A) $8.60$; (B) $14.33$; (C) $16.17$; (D) $57.33$.
E3. CSIR Part B GATE ST An RBD with $4$ treatments in $5$ blocks gives block mean square $26$ and error mean square $4$. The estimated efficiency of this design relative to a completely randomised design is closest to (A) $6.50$; (B) $2.38$; (C) $2.16$; (D) $1.37$.
E4. CSIR Part C Consider an experiment with $v$ treatments. Which of the following statements are true? (One or more options may be correct.) (A) In an RBD with $b$ blocks the error has $(v-1)(b-1)$ degrees of freedom. (B) The estimated efficiency of an RBD relative to a CRD exceeds $1$ if and only if the block mean square exceeds the error mean square. (C) Under the randomisation model of a CRD with unit–treatment additivity, the treatment means $\bar y_1,\dots,\bar y_v$ are uncorrelated. (D) In an RBD with random blocks and a random block × treatment interaction (unrestricted model), $\text{MSTr}/\text{MSE}$ has an exact $F$ distribution when there are no treatment effects.
E5. CSIR Part C Which of the following statements about Latin squares are true? (One or more options may be correct.) (A) A $2\times2$ Latin square design leaves no degrees of freedom for error. (B) There exist four mutually orthogonal Latin squares of order $5$. (C) There exists a pair of orthogonal Latin squares of order $6$. (D) There are exactly $6$ Latin squares of order $3$.
E6. CSIR Part C In a randomised block design with $v$ treatments and $b$ blocks one observation is missing and is replaced by its missing-plot estimate $\hat x$. Which of the following are true? (One or more options may be correct.) (A) The error degrees of freedom must be reduced by one. (B) The treatment sum of squares computed from the completed table is never smaller than the correct one. (C) $\hat x$ minimises the error sum of squares of the completed table. (D) The variance of the estimated difference between the treatment with the missing value and any other treatment is $2\sigma^2/b$.
Suggested plan — 3 lectures × 60 min.
- Principles and the CRD. Open with a badly designed trial (new variety on the good end of the field) and ask what analysis could rescue it — none. Definitions 8.10–8.12; Theorem 8.21 with its proof on the board (it reuses SRSWOR from §8.1, a nice link back); Exercise I.8 done live as a sanity check. Theorem 8.22 and Example 2 (ten minutes). Example 1 for computation.
- Randomised blocks. Definition 8.13, Theorem 8.23 (emphasise (1): block effects cancel); Example 3 including the one-way comparison at the end. Project Interactive 1 and switch between "blocks = rows" and "blocks = columns": students are surprised that badly placed blocks are worse than no blocks. Theorem 8.24 (proof of (3) only) and Example 4.
- Latin squares and missing plots. Definition 8.14, the orthogonality argument of Theorem 8.25, Example 5, Theorem 8.26. Then Theorem 8.27 as a general least-squares fact, Theorem 8.28 and Example 7 with Interactive 2 (move the missing cell around; the parabola keeps its shape). Finish with E4 and E6 of the Exam corner.
Discussion questions. (1) Randomisation makes the design unbiased "on average over allocations"; but the experiment is run only once. In what sense is the actual result protected? (2) Why does an LSD of order $4$ need only $16$ plots to control two nuisance factors that an RBD could control only with many more? What is assumed in exchange? (3) Is a missing value "missing at random" if the plot was flooded because it lay at the bottom of the slope?
Misconceptions to watch for. "Randomisation reduces error"; computing the RBD efficiency by pooling blocks with error; forgetting to remove an error df for a missing value; using the uncorrected treatment SS; believing that any arrangement with each letter once per row is a Latin square; treating subsamples as replicates; thinking random blocks require a different treatment test.
Suggested homework. I(2, 4, 6, 8), II(2, 4, 7, 8), III(1, 5, 6, 7), IV(2, 3, 5); Exam corner E2, E3, E5.
Research question. What shape and size should blocks have? Using a real uniformity trial — a field sown with a single variety and harvested plot by plot — measure how much different blocking schemes would have gained.
Why it matters. In 1910 W. B. Mercer and A. D. Hall harvested a wheat field at Rothamsted in $500$ small plots ($20$ rows by $25$ columns) with no treatments at all, to see how variable "identical" plots are. Their data showed strong, irregular spatial patterns, and they became the test bed for Fisher's ideas on blocking. The same question — how to group units so that within-group variation is small — arises when hospitals are used as blocks in multi-centre clinical trials, when production days are blocks in factory experiments, and in the design of crop variety trials worldwide today.
What to do.
- Obtain the Mercer–Hall grain yields (they are distributed with the R package
agridatasmercer.wheat.uniformity, and in several other open data collections). - Superimpose "dummy" experiments with $v=5$ treatments: (a) blocks of $5$ plots in a row, (b) blocks of $5$ plots in a column, (c) compact $1\times5$ versus $5\times1$ versus near-square groupings, (d) $5\times5$ Latin squares. Since all treatment effects are zero, the unit values are known exactly.
- For each scheme compute $S_u^2$ and the within-block variance $S_w^2$ (Theorem 8.24) and hence the true relative efficiency $S_u^2/S_w^2$; compare with $\widehat{\text{RE}}$ from a randomised dummy allocation.
- Repeat the randomisation $1000$ times for one scheme and check Theorem 8.24(1): the mean of the estimated treatment differences is $0$ and their variance is $2S_w^2/b$.
Tools: R or Python (pandas, numpy); a heat map of the field helps to see the gradients.
Hand in: a one-page table of efficiencies for the schemes you tried, a heat map of the field, and a recommendation for block shape with a reason.
Going further. Read about "nearest-neighbour" and spatial analyses of field trials (Papadakis' method, which adjusts each plot by its neighbours' residuals) and about the modern practice of fitting spatial correlation models, which treat the field as a random surface instead of fixed blocks.
🧪 Self-check quiz
Q1. Which principle of experimental design is what makes an estimate of the experimental error possible?
Q2. A $5\times5$ Latin square design has how many error degrees of freedom?
Q3. In an RBD with $v$ treatments and $b$ blocks, the missing-plot estimate (available totals $T$, $B$, $G$) is
Q4. The estimated efficiency of an RBD relative to a CRD is greater than 1 exactly when
Q5. Under unit–treatment additivity, what does randomisation of a CRD guarantee without any normality assumption?