Skip to content

CRD, RBD and Latin Square Designs

CSIR NET design of experiments: principles, CRD, RBD and Latin square models, ANOVA tables, missing plot formulas, efficiency and orthogonal Latin squares.

  • 19 min read
  • Updated
On this page · 14 sections
  1. Prerequisites
  2. The Hook: The Field With a Slope
  3. The Three Principles
  4. Completely Randomized Design (CRD)
  5. Randomized Block Design (RBD)
  6. Latin Square Design (LSD)
  7. Choosing Among the Three
  8. Memory Tricks
  9. Don’t Confuse
  10. Common Mistakes
  11. Practice
  12. Quick Revision Table
  13. Key Takeaways
  14. What’s Next

Prerequisites

The Hook: The Field With a Slope

The Three Principles

PrincipleWhat it doesWithout it
Randomizationassigns treatments to units by chance; makes the error estimate valid and protects against unknown systematic effectsestimates can be biased by the experimenter’s choices
Replicationrepeats each treatment; provides an estimate of error and reduces \( V(\bar y_i) = \frac{\sigma^2}{r} \)no error estimate, no test
Local control (blocking)groups homogeneous units so that comparisons are made within groups; reduces errorheterogeneity of units inflates the error

Terms. An experimental unit (plot) is the smallest unit to which a treatment is applied independently. Experimental error is the variation among units treated alike. A treatment contrast is \( \sum c_i\tau_i \) with \( \sum c_i = 0 \); an elementary contrast is \( \tau_i - \tau_j \).

All three designs below use a linear model with additive effects and errors \( e \sim N(0, \sigma^2) \) independent; the F tests need normality, the estimates do not.

Completely Randomized Design (CRD)

\( t \) treatments, treatment \( i \) on \( r_i \) units chosen completely at random from \( n = \sum r_i \) homogeneous units.

$$y_{ij} = \mu + \tau_i + e_{ij}, \qquad i = 1, \dots, t, \quad j = 1, \dots, r_i, \qquad \sum r_i\tau_i = 0.$$

With \( G \) the grand total, \( T_i \) the treatment totals and \( CF = \frac{G^2}{n} \):

SourcedfSS\( E(\text{MS}) \)
Treatments\( t - 1 \)\( \sum\frac{T_i^2}{r_i} - CF \)\( \sigma^2 + \frac{\sum r_i\tau_i^2}{t - 1} \)
Error\( n - t \)by subtraction\( \sigma^2 \)
Total\( n - 1 \)\( \sum\sum y_{ij}^2 - CF \)

Test \( H_0: \tau_1 = \dots = \tau_t \) with \( F = \frac{MS_{Tr}}{MS_E} \sim F_{t-1, n-t} \) under \( H_0 \). The estimate of \( \tau_i - \tau_j \) is \( \bar y_i - \bar y_j \), with variance \( \sigma^2\left(\frac{1}{r_i} + \frac{1}{r_j}\right) \).

  • Strengths: any number of treatments and replicates; unequal replication allowed; maximum error df; simplest analysis; a lost plot causes no difficulty.
  • Weakness: all unit-to-unit variation goes into error, so it suits only homogeneous material (laboratory, greenhouse).

Worked example

Three diets with replicates 4, 3, 5: P \( = (20, 23, 22, 19) \), Q \( = (25, 27, 26) \), R \( = (18, 17, 21, 20, 19) \). Totals 84, 78, 95; \( G = 257 \), \( n = 12 \).

SourcedfSSMSF
Diets292.9246.4619.01
Error922.002.444
Total11114.92

The standard error of \( \bar y_P - \bar y_Q = 21 - 26 \) is \( \sqrt{2.444\left(\frac{1}{4} + \frac{1}{3}\right)} \approx 1.19 \).

Randomized Block Design (RBD)

\( t \) treatments, \( r \) blocks of \( t \) homogeneous units each; every treatment appears once in every block, randomized separately within each block.

$$y_{ij} = \mu + \tau_i + \beta_j + e_{ij}, \qquad \sum\tau_i = \sum\beta_j = 0.$$
SourcedfSS\( E(\text{MS}) \)
Blocks\( r - 1 \)\( \frac{1}{t}\sum B_j^2 - CF \)\( \sigma^2 + \frac{t\sum\beta_j^2}{r - 1} \)
Treatments\( t - 1 \)\( \frac{1}{r}\sum T_i^2 - CF \)\( \sigma^2 + \frac{r\sum\tau_i^2}{t - 1} \)
Error\( (r - 1)(t - 1) \)by subtraction\( \sigma^2 \)
Total\( rt - 1 \)\( \sum\sum y_{ij}^2 - CF \)

Because every treatment appears once in each block, treatments and blocks are orthogonal: the treatment SS is the same whether or not blocks are fitted, and \( \bar y_{i\cdot} - \bar y_{i'\cdot} \) estimates \( \tau_i - \tau_{i'} \) free of block effects, with variance \( \frac{2\sigma^2}{r} \).

The hook analysed

\( G = 161 \), \( CF = \frac{161^2}{12} = 2160.083 \), \( \sum y^2 = 2239 \).

SourcedfSSMSF
Blocks330.2510.08324.2
Fertilisers246.1723.08355.4
Error62.500.417
Total1178.92

Standard error of a difference of two fertiliser means: \( \sqrt{\frac{2(0.417)}{4}} \approx 0.456 \).

Efficiency of RBD relative to CRD

From an RBD’s own ANOVA, the error mean square that a CRD on the same units would have had is estimated by

$$s^2_{CRD} = \frac{(r - 1)M_B + r(t - 1)M_E}{rt - 1},$$

where \( M_B \) and \( M_E \) are the block and error mean squares. The relative efficiency is \( \frac{s^2_{CRD}}{M_E} \). For the hook: \( \frac{3(10.083) + 8(0.417)}{11} = 3.053 \), and \( \frac{3.053}{0.417} \approx 7.3 \). (Some books multiply by a small correction for the different error df; it does not change the conclusion.)

Missing plot in an RBD

If one observation (treatment \( i \), block \( j \)) is lost, estimate it by the value that minimises the error SS:

$$\boxed{x = \frac{rB' + tT' - G'}{(r - 1)(t - 1)}}$$

where \( B' \), \( T' \), \( G' \) are the totals of the available observations in that block, for that treatment, and overall. Then:

  1. Insert \( x \) and run the usual ANOVA.
  2. Subtract 1 from the error df (and from the total df).
  3. The treatment SS so obtained is biased upward; the standard correction subtracts \( \frac{[B' - (t - 1)x]^2}{t(t - 1)} \) from it.
  4. The variance of the difference between the mean of the affected treatment and another treatment mean becomes \( \sigma^2\left[\frac{2}{r} + \frac{t}{r(r - 1)(t - 1)}\right] \).

Example. Lose B in block 3 of the hook (true value 13). \( B' = 11 + 9 = 20 \), \( T' = 16 + 17 + 18 = 51 \), \( G' = 148 \):

$$x = \frac{4(20) + 3(51) - 148}{3 \times 2} = \frac{85}{6} \approx 14.17.$$

With two missing plots, solve the two corresponding equations simultaneously, or iterate (guess one, compute the other, repeat).

Latin Square Design (LSD)

When heterogeneity runs in two directions (rows and columns of a field, or subjects and periods), arrange \( m \) treatments in an \( m \times m \) square so that each treatment appears exactly once in each row and once in each column.

$$y_{ijk} = \mu + \rho_i + \gamma_j + \tau_k + e_{ijk},$$

where \( k = k(i, j) \) is the treatment in cell \( (i, j) \); only \( m^2 \) of the \( m^3 \) combinations are observed.

SourcedfSS
Rows\( m - 1 \)\( \frac{1}{m}\sum R_i^2 - CF \)
Columns\( m - 1 \)\( \frac{1}{m}\sum C_j^2 - CF \)
Treatments\( m - 1 \)\( \frac{1}{m}\sum T_k^2 - CF \)
Error\( (m - 1)(m - 2) \)by subtraction
Total\( m^2 - 1 \)

Rows, columns and treatments are mutually orthogonal, so each SS is computed from its own totals. Error df: 0 for \( m = 2 \) (useless), 2 for \( m = 3 \), 6 for \( m = 4 \), 12 for \( m = 5 \). Small squares are often replicated to gain error df.

Missing plot in an LSD:

$$\boxed{x = \frac{m(R' + C' + T') - 2G'}{(m - 1)(m - 2)}}$$

with error df reduced by 1. For example, \( m = 4 \), \( R' = 30 \), \( C' = 28 \), \( T' = 26 \), \( G' = 150 \): \( x = \frac{4(84) - 300}{6} = 6 \).

Counting Latin squares

A standard (reduced) Latin square has its first row and first column in natural order. If \( L_m \) is the number of standard squares, the total number of Latin squares of order \( m \) is \( m!\,(m - 1)!\,L_m \) (permute all columns, then all rows but the first).

\( m \)Standard squares \( L_m \)All Latin squares
212
3112
44576
556161280

Randomization of an LSD: pick a standard square at random (for small \( m \)), then randomly permute rows, columns and treatment labels.

Orthogonal Latin squares and Graeco-Latin squares

Two Latin squares of order \( m \) are orthogonal if, when superimposed, each of the \( m^2 \) ordered pairs of symbols occurs exactly once. A Graeco-Latin square is such a superimposed pair; it controls three nuisance factors (rows, columns, Greek letters), leaving error df \( (m - 1)(m - 3) \).

  • There are at most \( m - 1 \) mutually orthogonal Latin squares (MOLS) of order \( m \).
  • For \( m \) a prime power, a complete set of \( m - 1 \) MOLS exists (built from the finite field \( GF(m) \): square \( a \) has entry \( ai + j \) in cell \( (i, j) \), \( a \ne 0 \)).
  • No pair of orthogonal Latin squares exists for \( m = 2 \) or \( m = 6 \); for every other \( m \ge 3 \) a pair exists.

Choosing Among the Three

graph TD
    A[Experimental units] --> B{Are units homogeneous?}
    B -->|Yes| C[CRD: maximum error df, simplest]
    B -->|No| D{Heterogeneity in how many directions?}
    D -->|One| E[RBD: blocks of t alike units]
    D -->|Two| F[Latin square: rows and columns, t = m]
    D -->|Blocks smaller than t| G[Incomplete block design, BIBD]
CRDRBDLSD
Nuisance factors removednoneone (blocks)two (rows, columns)
Restrictionnoneeach treatment once per blockeach treatment once per row and per column
Error df\( n - t \)\( (r - 1)(t - 1) \)\( (m - 1)(m - 2) \)
Number of treatmentsanyany (block size \( t \) must be feasible)\( m \) (equals rows = columns = replicates)
Missing plotno special treatment\( \frac{rB' + tT' - G'}{(r-1)(t-1)} \)\( \frac{m(R' + C' + T') - 2G'}{(m-1)(m-2)} \)

Memory Tricks

Don’t Confuse

ThisvsThat
Replication (repeats of a treatment)blocks (groups of alike units); in an RBD, number of blocks = replicates, but the ideas differ
RBD error df \( (r - 1)(t - 1) \)CRD error df \( t(r - 1) \) for the same \( rt \) units
Missing plot: error df reduced by 1treatment df unchanged
Standard (reduced) Latin squares: 4 of order 4all Latin squares of order 4: 576
Latin square: two-way eliminationGraeco-Latin square: three-way elimination, error df \( (m - 1)(m - 3) \)
Orthogonal Latin squares: every ordered pair onceorthogonal block design: \( n_{ij} = \frac{r_ik_j}{n} \) (next page)

Common Mistakes

Practice

All questions below are original practice questions written for this page. They are not from any CSIR NET paper. In Level 3, as in Part C, one or more options may be correct.

Level 1: Foundation

  1. An RBD has 5 treatments in 4 blocks. Give the error df.
  2. Give the error df of a \( 5 \times 5 \) Latin square, and of a \( 5 \times 5 \) Graeco-Latin square.
  3. Name the three basic principles of experimental design and the purpose of each.
Answers
  1. \( (4 - 1)(5 - 1) = 12 \).
  2. Latin: \( (4)(3) = 12 \). Graeco-Latin: \( (4)(2) = 8 \).
  3. Randomization (validity of the error estimate, protection against bias), replication (estimate of error, precision), local control (reduction of error by grouping alike units).

Level 2: Part B Style

Practice Q1. In a CRD with three treatments replicated 4, 3 and 5 times, the treatment SS is 92.92 and the total SS is 114.92. The F statistic for treatments is about

(a) 4.22 (b) 19.01 (c) 21.12 (d) 38.01

Answer: (b)

Error SS \( = 114.92 - 92.92 = 22.00 \) on \( 12 - 3 = 9 \) df: \( MS_E = 2.444 \). \( MS_{Tr} = \frac{92.92}{2} = 46.46 \). \( F = \frac{46.46}{2.444} = 19.01 \). Option (a) is the ratio of the sums of squares; (c) uses 10 error df; (d) forgets to divide the treatment SS by its 2 df.

Practice Q2. In an RBD with 3 treatments and 4 blocks, one observation is missing. For the affected block, treatment and the whole experiment, the totals of available observations are 20, 51 and 148. The missing value is estimated as

(a) 12.33 (b) 14.17 (c) 15.50 (d) 19.33

Answer: (b)

\( x = \frac{rB' + tT' - G'}{(r - 1)(t - 1)} = \frac{4(20) + 3(51) - 148}{6} = \frac{85}{6} = 14.17 \). Option (d) swaps \( r \) and \( t \): \( \frac{3(20) + 4(51) - 148}{6} = \frac{116}{6} = 19.33 \).

Practice Q3. In a \( 4 \times 4 \) Latin square one value is missing. The available totals for its row, column and treatment are 30, 28 and 26, and the available grand total is 150. The estimated missing value is

(a) 4 (b) 6 (c) 8 (d) 14

Answer: (b)

\( x = \frac{m(R' + C' + T') - 2G'}{(m - 1)(m - 2)} = \frac{4(84) - 300}{3 \times 2} = \frac{36}{6} = 6 \).

Practice Q4. The number of distinct Latin squares of order 4 (on four fixed symbols) is

(a) 4 (b) 24 (c) 144 (d) 576

Answer: (d)

There are 4 standard (reduced) squares of order 4; each gives \( 4! \times 3! = 144 \) squares by permuting all columns and the last three rows. Total \( 4 \times 144 = 576 \). Option (a) counts only standard squares; (c) forgets the factor 4.

Level 3: Part C Style

Practice Q5. For the RBD of the hook (3 fertilisers, 4 blocks, data in the table), which statements are true?

(a) The error mean square is about 0.42 on 6 df. (b) The F ratio for fertilisers is about 55.4 on \( (2, 6) \) df. (c) The estimated efficiency of this RBD relative to a CRD on the same units is about 7.3. (d) The standard error of the difference of two fertiliser means is about 0.65.

Answer: (a), (b) and (c)
  • (a) True. \( SS_E = 78.92 - 46.17 - 30.25 = 2.50 \); \( df = 3 \times 2 = 6 \); \( MS_E = 0.417 \).
  • (b) True. \( \frac{46.17/2}{0.417} = \frac{23.08}{0.417} = 55.4 \).
  • (c) True. \( s^2_{CRD} = \frac{3(10.083) + 8(0.417)}{11} = 3.053 \); \( \frac{3.053}{0.417} = 7.33 \).
  • (d) False. \( \sqrt{\frac{2 \times 0.417}{4}} = 0.456 \). The value 0.65 is \( \sqrt{0.417} \), the SE of a single plot.

Practice Q6. Which statements about Latin squares are true?

(a) There are at most \( m - 1 \) mutually orthogonal Latin squares of order \( m \). (b) A complete set of 6 mutually orthogonal Latin squares of order 7 exists. (c) A Graeco-Latin square of order 6 exists. (d) In a \( 4 \times 4 \) Graeco-Latin square design, the error has 3 df.

Answer: (a), (b) and (d)
  • (a) True. Standard bound: after normalising the first row of every square, the cells \( (2, 1) \) of the squares must carry distinct symbols different from the one in cell \( (1, 1) \), leaving at most \( m - 1 \) choices.
  • (b) True. 7 is prime; \( L_a(i, j) = ai + j \pmod 7 \), \( a = 1, \dots, 6 \).
  • (c) False. There is no pair of orthogonal Latin squares of order 6.
  • (d) True. \( 15 - 4 \times 3 = 3 = (m - 1)(m - 3) \).

Practice Q7. A CRD has \( t \) treatments with \( r_i \) replicates, \( n = \sum r_i \), model \( y_{ij} = \mu + \tau_i + e_{ij} \) with \( \sum r_i\tau_i = 0 \). Which statements are true?

(a) \( E(MS_{Tr}) = \sigma^2 + \frac{\sum r_i\tau_i^2}{t - 1} \). (b) \( \bar y_1 - \bar y_2 \) has variance \( \frac{2\sigma^2}{\bar r} \), where \( \bar r \) is the average replication. (c) The contrast \( 2\tau_1 - \tau_2 - \tau_3 \) is estimable. (d) The error df is \( n - t \), whether or not the replications are equal.

Answer: (a), (c) and (d)
  • (a) True. Standard one-way result (with the weighted side condition).
  • (b) False. It is \( \sigma^2\left(\frac{1}{r_1} + \frac{1}{r_2}\right) \); this equals \( \frac{2\sigma^2}{r} \) only when \( r_1 = r_2 = r \).
  • (c) True. Every treatment contrast is estimable in a CRD; its BLUE is \( 2\bar y_1 - \bar y_2 - \bar y_3 \).
  • (d) True. \( n - 1 \) total minus \( t - 1 \) for treatments.

Quick Revision Table

IdeaFact
Principlesrandomization (validity), replication (error estimate, precision), local control (error reduction)
CRD\( y_{ij} = \mu + \tau_i + e_{ij} \); error df \( n - t \); \( V(\bar y_i - \bar y_j) = \sigma^2(\frac{1}{r_i} + \frac{1}{r_j}) \)
RBDblocks \( r - 1 \), treatments \( t - 1 \), error \( (r-1)(t-1) \); \( V(\bar y_i - \bar y_j) = \frac{2\sigma^2}{r} \)
RBD missing\( x = \frac{rB' + tT' - G'}{(r-1)(t-1)} \); error df \( -1 \)
RE of RBD vs CRD\( \frac{(r-1)M_B + r(t-1)M_E}{(rt-1)M_E} \)
LSDrows, columns, treatments \( m - 1 \) each; error \( (m-1)(m-2) \)
LSD missing\( x = \frac{m(R' + C' + T') - 2G'}{(m-1)(m-2)} \)
Counting\( m!(m-1)!L_m \); \( L_3 = 1, L_4 = 4, L_5 = 56 \); 12, 576, 161280
MOLSat most \( m - 1 \); \( m - 1 \) for prime powers; none for \( m = 2, 6 \)
Graeco-Latinerror df \( (m-1)(m-3) \)

Key Takeaways

What’s Next